Announce: "Using workflow-creator to design/audit/improve a structured workflow."

Load workflow-creator's own constraints (auto-discovered + applies-to-filtered — surfaces the wc-* behavioral rules at load time, complementing the wc-constraint-check.py post-edit hook):

!uv run python3 ${CLAUDE_SKILL_DIR}/../../scripts/load-constraints.py workflow-creator

Detect mode from user request, then follow the corresponding process below:

Mode 1 (Create) — "create/design a workflow", "break a task into phases"
Mode 2 (Audit) — "audit/score a workflow"
Mode 3 (Improve) — "improve a workflow", audit-fix loop, "migrate a fan-out phase to an ultracode workflow" (a migration is an improvement — see ${CLAUDE_SKILL_DIR}/references/dynamic-workflow-migration.md)

Note on workflow-creator's Structure:

workflow-creator is a meta-tool that CREATES workflows. It is exempt from certain requirements it enforces on workflows it creates:

Two entry points: workflow-creator has one entry with mode detection (not a multi-phase workflow). Workflows it creates MUST have two entry points.
Single responsibility per phase: workflow-creator has 3 modes (toolkit, not workflow). Workflows it creates MUST have single-responsibility phases.
Skill Dependencies (cross-file phase chaining): workflow-creator is a single SKILL.md, so it has no next-phase Read() chain. Its structural equivalent is stronger: the wc-step-gate-guard.py hook + the STATE.md step-chain enforce step ordering at the tool-call layer (a skipped step is blocked, not merely un-chained). Workflows it creates with multiple phase files MUST still wire phase-to-phase Read() transitions.
Iteration topology labels (P09): workflow-creator's own steps are a fixed sequence, not a per-phase topology menu, so they carry no [one-shot|serial|parallel] label. Workflows it creates MUST assign a topology per phase.

This document defines the PROCESS for creating workflows. The workflows created by this process must follow all principles from PHILOSOPHY.md.

The mode flowcharts below ARE the authoritative spec. Each mode (Create, Audit, Improve) opens with an ASCII step/phase diagram. If the prose for a step ever conflicts with its mode's flowchart, the flowchart wins — treat a diagram-violating shortcut as a process error, not a "minor deviation from documentation."

Startup: State Check

Before detecting mode, check for existing workflow-creator state:

IDENTIFY: Run Glob(".planning/wc/*/HANDOFF.md") and Glob(".planning/wc/*/STATE.md")
READ: If any HANDOFF.md found → read it. If STATE.md found without HANDOFF.md → read STATE.md to determine last completed step.
VERIFY: If HANDOFF.md exists, confirm the recorded mode/step match what STATE.md shows. If they conflict, trust STATE.md (it's hook-enforced).
DECIDE: If resumable state found → offer to resume from recorded position (skip mode detection). If not found → proceed with mode detection below. [checkpoint: decision]

Determining {name}: The {name} in all state file paths is the target workflow name (e.g., dev, ds, writing, teaching). For Mode 1, use the proposed workflow name from the interview. For Modes 2-3, use the workflow being audited/improved.

Why .planning/wc/: workflow-creator's state files must NOT conflict with the target project's .planning/ state files (SPEC.md, PLAN.md, STATE.md, etc.). The wc/ subdirectory isolates workflow-creator's meta-state from the workflow state it's auditing or creating.

Namespace by target workflow: Each workflow-creator invocation operates on a specific target workflow. State files go in .planning/wc/{workflow-name}/ to prevent clashes when auditing/improving multiple workflows in parallel (e.g., parallel companions auditing dev, ds, writing simultaneously).

.planning/wc/
├── dev/                    → audit/improve state for dev workflow
│   ├── STATE.md
│   ├── AUDIT.md
│   └── SCORES.md
├── ds/                     → audit/improve state for ds workflow
│   ├── STATE.md
│   ├── AUDIT.md
│   └── SCORES.md
└── writing/                → audit/improve state for writing workflow
    ├── STATE.md
    └── AUDIT.md

For Mode 1 (create), use the proposed workflow name: .planning/wc/{new-workflow-name}/.

Standard workflow-creator state files:

| File | Purpose | Created By | |------|---------|-----------| | .planning/wc/{name}/STATE.md | Current mode + step | All modes at startup | | .planning/wc/{name}/INTERVIEW.md | Captured interview answers | Mode 1 Step 2 | | .planning/wc/{name}/DESIGN.md | Phase decomposition decisions | Mode 1 Step 3 | | .planning/wc/{name}/AUDIT.md | Audit findings and scores | Mode 2 Step 4, Mode 3 Phase A | | .planning/wc/{name}/SCORES.md | Score history across iterations | Mode 3 Phase A | | .planning/wc/{name}/VALIDATION.md | Maps each WC-NN requirement → verification evidence (which gate/audit confirms it) + scope tag (v1/v2/out-of-scope) | Mode 1 Step 7, Mode 2 Step 4 | | .planning/wc/{name}/LEARNINGS.md | Log of what the user attended to / changed at present-to-user checkpoints (Step 6/7) — the observe→record→offer loop wc prescribes for created workflows, applied to itself | Mode 1 Step 6-7 | | .planning/wc/{name}/HANDOFF.md | Session resume context | Any mode on context exhaustion |

wc HANDOFF.md template (workflow-creator's own handoff — the same structured format it mandates for the workflows it creates, applied to itself, so a resuming session starts immediately without re-discovery):

---
mode: create | audit | improve
step: <current step or phase>
status: paused
target: <workflow name>
context_remaining: <e.g. 24%>
last_updated: <ISO8601 — pass in via args; do not invent>
---
## Current State
<what is in progress right now>
## Completed Work
<steps done + key artifacts written>
## Remaining Work
<steps/phases left>
## Decisions Made / Rejected Approaches
<so the resume doesn't relitigate>
## Next Action
<specific enough to start immediately — not "continue">

Mode 1: Create New Workflow

Step 1: Philosophy ──→ Step 2: Interview ──→ Step 3/3b: Decomposition + Artifact Gates
  [auto]                [pause: interview]     [auto]
    │                       │                      │
    ▼                       ▼                      ▼
  STATE.md updated        INTERVIEW.md           DESIGN.md
                          STATE.md updated       STATE.md updated
                                                     │
Step 7: Self-Audit ◄── Step 6: Generate ◄── Step 5: Entry Points ◄── Step 4/4b: Enforcement
  [decision: present]    [decision: present]   [auto]                    [auto]
    │                       │                      │                        │
    ▼                       ▼                      ▼                        ▼
  AUDIT.md via subagent   Skill files written    STATE.md updated         STATE.md updated
  Score ≥ 8.0? ──NO──→ Fix ──→ Re-audit (max 3)
    │ YES
    ▼
  Present to user

<EXTREMELY-IMPORTANT> **NO PAUSE BETWEEN STEPS.** After completing each step, immediately start the next. Do NOT ask "should I continue?", do NOT summarize what you just did, do NOT wait for confirmation — pause ONLY where explicitly required (Step 6: present files; Step 7: present audit results). Pausing between steps is procrastination disguised as courtesy: it strands the workflow and hands the user a management burden they should never have to carry. </EXTREMELY-IMPORTANT> <EXTREMELY-IMPORTANT> **Enforcement architecture:** Step transitions are hook-enforced via `wc-step-gate-guard.py`: - **Layer 2 (step-chain):** Writing `step: N` to STATE.md is BLOCKED unless `step: N-1` shows `status: completed`. This fires for ALL modes (create, audit, improve). - **Layer 1 (file-path gates):** Writing INTERVIEW.md, DESIGN.md, AUDIT.md, and skill/constraint files is BLOCKED unless the prerequisite step is completed. Mode 1 only.

Every step below has a STATE.md YAML template. You MUST write this template to STATE.md BEFORE advancing — the hook enforces the chain.

If you catch yourself thinking "I can skip the STATE.md update" — STOP. The STATE.md update IS the gate artifact. The hook will BLOCK subsequent writes without it. </EXTREMELY-IMPORTANT>

Step 1: Ground in Philosophy

Discover and read PHILOSOPHY.md:Read ${CLAUDE_SKILL_DIR}/../../PHILOSOPHY.md and follow its instructions. You MUST read this file before proceeding. No claiming you "remember" it. Every workflow must address: phased decomposition, gates (deterministic or judgment-based), independent verification, artifact review, iteration strategy, and two entry points.

Gate: Philosophy Loaded [checkpoint: human-verify, auto-advanceable]

Verify PHILOSOPHY.md was read
Check that your response references: phased decomposition, gates, independent verification, artifact review, iteration strategy, two entry points
If you cannot explain these principles, re-read PHILOSOPHY.md

After verifying Philosophy is loaded, write initial state:

mkdir -p .planning/wc/{name} && cat > .planning/wc/{name}/STATE.md << 'EOF'
---
mode: create
step: 1-philosophy
status: completed
implements: [WC-01]
requires: [PHILOSOPHY.md]
provides: []
affects: [.planning/wc/{name}/STATE.md]
key-files: {read: [PHILOSOPHY.md]}
one-liner: "Philosophy grounded — phased decomposition, gates, independent verification, artifact review, iteration, two entry points confirmed."
---
Philosophy loaded. Proceeding to interview.
EOF

Proceed to Step 2. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 2: Interview

Context check: The interview is interactive and may require multiple exchanges. Before proceeding:

If context is low (≤35% remaining), write .planning/wc/{name}/HANDOFF.md with philosophy status and current progress. Pause.
If context is critical (≤25% remaining), write HANDOFF.md immediately.

Red Flags — STOP: inferring the domain from context instead of asking · skipping questions because they "seem obvious" or the user "seems busy" · closing the interview before all 6 answers are concrete (a shallow interview makes every downstream phase wrong — you'll design for an imagined domain, not the real one). Batch all 6 into one AskUserQuestion; don't drip them.

Use AskUserQuestion to understand the domain:

What kind of work? (code, data, writing, research, other)
What's the deliverable? (working feature, analysis report, polished document, etc.)
What are the common failure modes? (skipping tests, shallow analysis, weak arguments, etc.)
When does drift happen? (implementation without design, conclusions without evidence, etc.)
How should iteration work? (one-shot with verification, serial hypothesis testing, parallel exploration, agent team review)
What does verification look like? (running tests, checking output exists, reviewing summary artifact — define concretely so "verification" can't become investigation)

Gate: Interview Complete [checkpoint: human-verify, auto-advanceable]

Verify AskUserQuestion was called
Check that answers to all 6 questions are present
If interview incomplete, ask remaining questions

After verifying Interview is complete, persist answers and update state:

Write .planning/wc/{name}/INTERVIEW.md with all 6 answers in structured format:

---
workflow_name: [proposed name]
domain: [code/data/writing/research/other]
---
## Answers
1. **Work type:** ...
2. **Deliverable:** ...
3. **Failure modes:** ...
4. **Drift points:** ...
5. **Iteration style:** ...
6. **Verification:** ...

Update .planning/wc/{name}/STATE.md:

step: 2-interview
status: completed
implements: [WC-02]
requires: [PHILOSOPHY.md]
provides: [INTERVIEW.md]
affects: [.planning/wc/{name}/INTERVIEW.md, .planning/wc/{name}/STATE.md]
key-files: {created: [.planning/wc/{name}/INTERVIEW.md]}
one-liner: "Captured domain, deliverables, failure modes, drift points, iteration style, and verification approach."

INTERVIEW.md Review Gate

Before proceeding to decomposition, verify the interview capture is complete and unambiguous. Dispatch a lightweight reviewer:

Agent(
  subagent_type="general-purpose",
  description="Review INTERVIEW.md completeness",
  allowed_tools=["Read", "Grep", "Glob"],
  prompt="""Read .planning/wc/{name}/INTERVIEW.md.

Check against the 6 required interview questions:
1. Work type  2. Deliverable  3. Failure modes  4. Drift points  5. Iteration style  6. Verification

For each question, verify the answer is:
- Present (not missing or placeholder)
- Specific enough for decomposition (not "TBD" or "various")
- Consistent with the stated domain

Report: APPROVED if all 6 are adequate, or list specific gaps.
Do NOT edit the file — report only."""
)

Gate: Interview Reviewed [checkpoint: human-verify, auto-advanceable]

If reviewer reports APPROVED → write the structural marker .planning/wc/{name}/INTERVIEW_REVIEWED.md (frontmatter status: APPROVED, timestamp, 1-line summary of what was verified), then proceed. This is hook-enforced: wc-step-gate-guard.py (Layer 3) BLOCKS the STATE.md write for step: 3-decomposition unless this marker exists with status: APPROVED — the same structural-gate-artifact pattern this skill mandates for the workflows it creates, applied to itself, not advisory trust.
If reviewer reports gaps → drive convergence via /goal Interview reviewer returns APPROVED on .planning/wc/{name}/INTERVIEW.md. Stop after 5 turns. Each turn: ask remaining questions, update INTERVIEW.md, re-dispatch the reviewer, end turn

Proceed to Step 3. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 3: Propose Phase Decomposition

Gate prerequisite (structural): Refuse to start Step 3 unless .planning/wc/{name}/INTERVIEW_REVIEWED.md exists with status: APPROVED. If it is missing, return to the INTERVIEW.md Review Gate — decomposition built on an unreviewed interview inherits its gaps.

Context check: Decomposition and artifact gate design (Steps 3-3b) produce DESIGN.md — the recoverable artifact for enforcement generation. Before proceeding:

If context is low (≤35% remaining), write .planning/wc/{name}/HANDOFF.md with interview answers (from INTERVIEW.md) and current progress. Pause.
If context is critical (≤25% remaining), write HANDOFF.md immediately.

Design phases where each phase has:

Name - verb-noun (e.g., explore-codebase, design-approach)
Responsibility - ONE question this phase answers (single responsibility principle)
Gate condition - verifiable exit criterion (file exists, test passes, artifact contains X)
Gate artifact - the concrete file the producing phase writes and the consuming phase checks (see Structural Gate Artifacts below)
Enforcement needs - high/medium/low based on drift risk

Ultracode-workflow check: For any phase that is a fan-out over a known list — either read-only review (one reviewer per item → gate/findings) OR write/transform (one write-agent per item that creates/transforms from a fixed spec — codemod, migration, per-item spec-driven generation, worktree-isolated) — decide whether to implement it as a Claude Code ultracode workflow rather than in-skill agent dispatch. Workflows are NOT read-only; the docs' flagship case is a 500-file write migration. Read ${CLAUDE_SKILL_DIR}/references/dynamic-workflow-migration.md for the decision rubric, the hybrid split (workflow = deterministic fan-out read-or-write; skill keeps creative drafting + /goal + R4 + user input), and script conventions. Keep only creative/judgment generation, user-approval, and /goal loops conversational — mechanical/spec-driven per-item creation belongs in a transform workflow.

Structural Gate Artifacts

<EXTREMELY-IMPORTANT> **Every mandatory gate between phases MUST be enforced by a concrete artifact — not instructional text.**

Advisory gates ("you must run X before proceeding", "prerequisite: Y complete") are unenforceable. The agent that skips a gate doesn't check whether it was supposed to run. The consuming phase must structurally refuse to start without the artifact.

The pattern:

Producing phase writes a marker file (e.g., .planning/PHASE_REVIEWED.md) with status frontmatter
Consuming phase checks for the file at startup and REFUSES to proceed without it
The marker file includes: status: APPROVED/COMPLETED, timestamp, and summary of what was verified

Example (from dev workflow):

dev-plan-reviewer writes → .planning/PLAN_REVIEWED.md (status: APPROVED)
dev-implement checks   → file exists AND status == APPROVED, else REFUSE to start

Naming convention: .planning/{PHASE_NAME}_{ACTION}.md (e.g., SPEC_REVIEWED.md, DESIGN_APPROVED.md, EXPLORATION_COMPLETE.md)

Red Flags — STOP if you catch yourself thinking:

"The skill says 'must' so it will be followed" → STOP. Advisory text is not enforcement. The agent that would skip the gate is the same agent reading the advisory text.
"The entry point chains phases automatically so gates can't be skipped" → STOP. Users can invoke any phase skill directly. Mid-point entry bypasses the chain.
"Adding marker files is overhead" → STOP. The dev workflow shipped advisory-only gates for months. The bug was only caught by a meta-audit, not by the workflow itself.

| Gate Type | Enforcement | Can Be Bypassed? | Strength | |-----------|-------------|-------------------|----------| | Advisory | "You must run X first" | Yes — agent rationalizes past it | Weakest | | Artifact check | Instructional text checks for file at startup | Mostly no — but can be skipped under context pressure | Medium | | Hook-enforced | PreToolUse blocks tool calls until artifact exists | No — Claude Code blocks before action, no rationalization possible | Strongest |

Design every inter-phase gate as hook-enforced. Artifact checks are fallback. Never advisory-only.

Hook-Enforced Gates (Preferred Pattern)

Use the generic phase-gate-guard.py hook to enforce gate artifacts at runtime. The hook blocks Write/Edit/Agent tools until the required artifact exists with the correct status.

Frontmatter pattern for consuming phase:

hooks:
  PreToolUse:
    - matcher: "Write|Edit|Agent"
      hooks:
        - type: command
          command: >-
            GATE_ARTIFACT=.planning/PLAN_REVIEWED.md
            GATE_STATUS=APPROVED
            GATE_DESCRIPTION="Plan review"
            GATE_REMEDY="Return to dev-design and run dev-plan-reviewer"
            uv run python3 ${CLAUDE_PLUGIN_ROOT}/hooks/phase-gate-guard.py

How it works:

Producing phase writes .planning/X_REVIEWED.md with status: APPROVED frontmatter (unchanged)
Consuming phase declares a PreToolUse hook that checks for the artifact
Claude Code blocks Write/Edit/Agent calls until the artifact exists
Writes to .planning/ and .claude/ are always allowed (the phase can still write state files)

Why hooks > artifact checks in instructions:

Instructions can be compressed away during context compaction
Claude can rationalize "the file probably exists, I'll check later"
Hooks fire on EVERY tool call — no escape, no rationalization, no context dependency

When designing gates for a new workflow, generate BOTH:

The artifact (producing phase writes it)
The hook (consuming phase declares it in frontmatter) </EXTREMELY-IMPORTANT>

Critical: Each phase must have exactly ONE responsibility. If a phase does two things, split it into two phases. Phased decomposition means clean boundaries between concerns.

Present 2-3 topologies to the user:

Linear - phase 1 → phase 2 → ... → phase N (best for predictable work)
Branching - routing based on input type (best for varied work like writing)
Iterative - phases with loops (best for exploratory work like DS)

Iteration Topology

Based on the interview answer about iteration, assign each phase an iteration strategy:

| Strategy | When to Use | Implementation | |----------|------------|----------------| | One-shot + verify | Clear specs, low ambiguity | Single subagent, run tests, move on | | Serial hypothesis | Debugging, root cause analysis | Fresh subagent per iteration, HYPOTHESES.md as memory, progress-gated escalation | | Parallel exploration | Multiple valid approaches, robustness checking | Spawn N subagents simultaneously, converge findings in state file | | Agent team | Output needs multi-faceted review | Specialized reviewer subagents in parallel (e.g., copy + critic + fact-check), consolidate in REVIEW.md |

CRITICAL — Flat Dispatch Only: For "Parallel exploration" and "Agent team" strategies, the orchestrator (phase skill or main chat) MUST spawn all agents directly. Never design an intermediate "coordinator" or "dispatcher" agent that spawns sub-agents on its own. See Iron Law of Flat Dispatch.

Exit conditions by strategy:

| Strategy | Exit Gate | Escalate When | |----------|-----------|---------------| | One-shot | Test passes | Test fails after fix attempt | | Serial | New findings stop emerging | 3+ consecutive failures, repeated hypotheses | | Parallel | Findings converge | Results contradictory, no convergence | | Agent team | Reviewers converge | Unresolvable disagreement on direction |

Key principle: The agent never declares its own completion. Tests pass, findings converge, or the human approves.

Verification Depth

When designing verification phases, ensure they check all 4 levels — not just existence:

| Level | Question | Gate Fails If... | |-------|----------|-----------------| | 1. Exists | Is the deliverable physically present? | File/function/test missing | | 2. Substantive | Is it real, not a stub? | Body is pass, TODO, placeholder, or trivial | | 3. Wired | Is it connected to the system? | Defined but never imported, called, or routed | | 4. Functional | Does it actually work? | Tests fail, feature errors at runtime |

Verification gates that only check Level 1 ("file exists") are theater. Design gates that verify through Level 4 where possible.

Test Gap Validation Phase

Workflows with implementation phases should include a validation phase between implement and review. This phase maps every requirement from the spec to test coverage, classifying each as COVERED / PARTIAL / MISSING, and fills gaps before review begins.

Why: Implementation subagents write tests per-task, but gaps hide between tasks. A dedicated validation pass catches requirements that no single task covered.

Phase design:

Read requirements from spec
Scan existing tests and map each requirement to coverage
Classify: COVERED / PARTIAL / MISSING
Fill gaps (write new tests, not implementation fixes)
Produce VALIDATION.md with the full coverage map

Gate condition: VALIDATION.md exists with status validated — all requirements COVERED, all tests passing.

Checkpoint Types

Not all gates are the same. GSD distinguishes three checkpoint types with dramatically different frequencies:

| Type | Frequency | Description | What Happens | |------|-----------|-------------|-------------| | human-verify | ~90% | Agent did the work, human confirms | Review and approve (auto-advanceable) | | decision | ~9% | Human chooses direction from options | Select from options with pros/cons | | human-action | ~1% | Auth gates, 2FA, physical access | Human performs truly manual step |

When designing gates, classify each one. Most gates are human-verify — the agent can auto-advance them in autonomous mode. Only decision (choose between approaches) and human-action (credentials, physical access) require genuine human pause.

Golden rule: If the agent CAN automate it, the agent MUST automate it. human-action is reserved for things genuinely impossible to automate.

Why this matters: Without checkpoint classification, every gate pauses for human input. Workflows become unusable in autonomous/overnight mode because they stop at every human-verify checkpoint that could have been auto-approved.

Context Monitoring

Long workflows must plan for context exhaustion. Without monitoring, agents start complex work when context is nearly full, produce degraded output, and lose in-flight state.

Requirements for workflows:

Graceful degradation — phases should check context availability before starting expensive work
Handoff trigger — when context is low, trigger .planning/HANDOFF.md creation instead of starting a new phase
Phase-aware warnings — implementation phases need more remaining context than exploration phases

Implementation pattern:

At phase entry, check if sufficient context remains for the phase's expected work
If context is low (≤35% remaining), write .planning/HANDOFF.md and pause rather than starting degraded work
If context is critical (≤25% remaining), immediately write .planning/HANDOFF.md — no new work

Standard thresholds:

| Level | Remaining Context | Action | |-------|------------------|--------| | Normal | >35% | Proceed normally | | Warning | 25-35% | Complete current task, then handoff | | Critical | ≤25% | Immediate handoff, no new tasks |

Why: Context exhaustion is the #1 cause of lost work in long workflows. An agent that starts a 10-task implementation phase with 20% context remaining will produce garbage for the last 5 tasks. Better to handoff cleanly and resume fresh.

Summary Frontmatter

Phase completions should produce structured YAML summaries for machine-readable context assembly. This enables automated resume, dependency analysis, and audit trails.

Phase SUMMARY.md format:

---
phase: explore-codebase
status: completed
duration: 12m
implements: [REQ-01, REQ-03]
requires: [SPEC.md]
provides: [EXPLORATION.md]
affects: [src/auth/, src/middleware/]
key-files:
  created: [tests/test_auth.py]
  modified: [src/auth/handler.py]
deviations: {r1: 1, r2: 0, r3: 1, r4: 0}
tags: [authentication, middleware]
---

One-liner: JWT auth exploration — identified 3 integration points and 2 missing test paths.

## Findings
...

Required fields:

phase, status — identification
implements — which requirement IDs this phase addressed
requires / provides — dependency graph between phases
affects — directories/files changed (for conflict detection)
key-files.created, key-files.modified — file tracking
deviations — R1-R4 counts from deviation rules

One-liner rule: Must be SUBSTANTIVE. Good: "JWT auth with refresh rotation using jose". Bad: "Phase complete" or "Implemented authentication".

Why: Without structured summaries, handoff and resume require re-reading all changed files. With frontmatter, the next session can reconstruct what happened from provides/affects fields without reading the full phase output.

Agent Tool Restrictions (READ-ONLY Verifiers)

Verification agents must be structurally prevented from modifying the work they verify. A verifier that can Write/Edit will "fix" issues it discovers, bypassing the plan-execute-verify cycle.

Implementation: Use allowed-tools frontmatter on verification/review agents:

---
name: code-reviewer
description: Reviews code for quality issues
allowed-tools:
  - Read
  - Grep
  - Glob
  - Bash(command_prefix:cat)
  - Bash(command_prefix:git log)
  - Bash(command_prefix:git diff)
---

Tool restriction tiers:

| Agent Role | Can Use | Cannot Use | |-----------|---------|------------| | Executor | Read, Write, Edit, Bash, Grep, Glob | — | | Verifier | Read, Grep, Glob, Bash (read-only commands) | Write, Edit, Bash (modifying commands) | | Researcher | Read, Grep, Glob, WebFetch, WebSearch | Write, Edit, Bash | | Auditor | Read, Grep, Glob | Write, Edit, Bash |

Why: Without tool restrictions, "independent verification" is a polite fiction. The verifier reads, finds a bug, fixes it in-place, and reports "all checks pass." The fix was never planned, never reviewed, and never tested. Tool restrictions make verification structurally honest.

Requirement Traceability

Requirements should have unique IDs that flow through the entire workflow — from spec through plan through implementation through verification.

Tracing chain:

SPEC.md assigns unique IDs per requirement (e.g., AUTH-01, AUTH-02, DATA-01)
PLAN.md tasks reference requirement IDs (implements: [AUTH-01, AUTH-02])
Phase summaries track which IDs were addressed (implements: [AUTH-01])
VALIDATION.md maps every ID to test evidence (COVERED / PARTIAL / MISSING)
Milestone audit checks all v1 requirements are satisfied before marking complete

ID format: CATEGORY-NN (e.g., AUTH-01, DATA-03, UI-12). Categories come from natural groupings in the spec.

Scope classification:

| Tag | Meaning | |-----|---------| | v1 | Must be complete for milestone | | v2 | Nice to have, defer if needed | | out-of-scope | Explicitly excluded |

Why: Without IDs, requirement-to-test mapping is fuzzy. "We tested authentication" doesn't tell you whether AUTH-01 (login), AUTH-02 (refresh tokens), and AUTH-03 (logout) are all covered. IDs make gaps visible and auditable.

Autonomous Phase Chaining

Workflows should support autonomous execution — chaining phases automatically without human intervention at every step.

Key mechanisms:

Smart Discuss — batch all ambiguities into one question instead of sequential asks. Present all grey areas at once for a single human response.
Dynamic phase re-read — after each phase completes, re-read the ROADMAP/PLAN to catch dynamically inserted phases (phases added during execution of an earlier phase).
Checkpoint-aware pausing — only pause at decision and human-action checkpoints; auto-advance human-verify checkpoints.
Blocker handling — when execution fails, offer: retry / skip / stop options.
Post-execution routing — based on verification status, route to: next phase / retry / human escalation.

Auto-advance mode: Auto-approves human-verify checkpoints, auto-selects first option for decision checkpoints. Only human-action pauses.

Why: Without autonomous chaining, the user must manually invoke each phase. A 7-phase workflow requires 7 manual interventions. With autonomous mode, the user kicks off the workflow and returns to find it complete (or paused at a genuine decision point).

Gate: Decomposition Complete [checkpoint: human-verify, auto-advanceable]

Every proposed phase has a single responsibility, a verifiable gate condition, a named gate artifact, and an iteration topology
The topology choice (linear/branching/iterative) presented to the user is recorded [checkpoint: decision]
The fan-out / ultracode-workflow check was applied to every phase
If any phase lacks a gate or has >1 responsibility, split/fix it before advancing

Update .planning/wc/{name}/STATE.md:

step: 3-decomposition
status: completed
implements: [WC-03]
requires: [PHILOSOPHY.md, INTERVIEW.md]
provides: [phase decomposition, gate conditions, enforcement needs]
affects: [.planning/wc/{name}/STATE.md]
one-liner: "Phases decomposed with single responsibilities, gate conditions, and enforcement needs."

Proceed to Step 3b. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 3b: Add Artifact Review Gates

Context check: Step 3b produces DESIGN.md — the recoverable artifact for enforcement generation. Before proceeding:

If context is low (≤35% remaining), write .planning/wc/{name}/HANDOFF.md with decomposition progress and current DESIGN.md draft. Pause.
If context is critical (≤25% remaining), write HANDOFF.md immediately.

Red Flags — STOP: marking an inter-phase gate "advisory-only" to move faster (advisory gates are the #1 audit gap — P03 — and you are designing the exact transition that will later fail) · designing a gate the consuming phase never structurally checks · assuming "the SKILL.md says must, so it will be followed" (the agent that skips the gate is the one reading that prose). Prefer hook-enforced gate artifacts; advisory is never acceptable for a mandatory gate.

For every phase that produces an artifact consumed by downstream phases, add an artifact review gate between the producing phase and the consuming phase.

Phase N produces ARTIFACT.md
  → Dispatch independent reviewer subagent
  → Reviewer checks: completeness, consistency, clarity, YAGNI, spec alignment
  → If ISSUES_FOUND → drive convergence via /goal pinned to reviewer APPROVED, 5-turn budget
  → If APPROVED → Phase N+1 consumes the artifact

Common artifact-producing phases: | Artifact | Typical Producer | Typical Consumer | |----------|-----------------|------------------| | Spec/requirements | Brainstorm | Explore, Design | | Plan/task list | Design | Implement | | VALIDATION.md | Validate (test gap) | Review | | Outline | Brainstorm | Draft | | Hypothesis list | Investigate | Test |

VALIDATION.md gates the transition from implement to review. Without it, review has no evidence that requirements were tested — it can only review what it sees, not what's missing. The validation phase produces this artifact; the review phase consumes it.

Chunking rule: If the artifact has >15 discrete items (tasks, requirements, sections), break into ordered chunks and review each separately.

Model tier guidance: Add to any phase that dispatches implementation subagents:

Mechanical tasks (1-2 files, clear spec) → cheapest capable model
Integration tasks (multi-file coordination) → standard model
Architecture/review tasks (design judgment) → most capable model

Gate: Artifact Review Gates Designed [checkpoint: human-verify, auto-advanceable]

Every artifact-producing phase has a review gate before the consuming phase
Reviewer is a fresh subagent (not self-review)
Fix-and-re-review loop is /goal-driven (5-turn budget; evaluator gates exit on reviewer APPROVED)
Chunking specified for large artifacts

After verifying Artifact Review Gates are designed, persist design decisions:

Write .planning/wc/{name}/DESIGN.md with phase decomposition, topology choice, iteration strategies, and artifact review gates. This is the recoverable artifact if context exhausts during enforcement generation.

Update .planning/wc/{name}/STATE.md:

step: 3b-artifact-review
status: completed
implements: [WC-03, WC-04]
requires: [PHILOSOPHY.md, INTERVIEW.md]
provides: [DESIGN.md]
affects: [.planning/wc/{name}/DESIGN.md]
key-files: {created: [.planning/wc/{name}/DESIGN.md]}
one-liner: "Phase decomposition, topology, iteration strategies, and artifact review gates finalized."

DESIGN.md Review Gate

Before proceeding to enforcement generation, verify the decomposition design is sound. Dispatch a lightweight reviewer:

Agent(
  subagent_type="general-purpose",
  description="Review DESIGN.md completeness",
  allowed_tools=["Read", "Grep", "Glob"],
  prompt="""Read .planning/wc/{name}/DESIGN.md.

Check against decomposition requirements:
1. Each phase has a single responsibility (one question answered)
2. Every phase has a gate condition (verifiable exit criterion)
3. Gate artifacts are specified (concrete files, not prose)
4. Iteration topology is assigned per phase
5. Artifact review gates exist between producing/consuming phases

Report: APPROVED if design is sound, or list specific gaps.
Do NOT edit the file — report only."""
)

Gate: Design Reviewed [checkpoint: human-verify, auto-advanceable]

If reviewer reports APPROVED → write the structural marker .planning/wc/{name}/DESIGN_REVIEWED.md (frontmatter status: APPROVED, timestamp, summary of what was verified), then proceed. This is hook-enforced: wc-step-gate-guard.py (Layer 3) BLOCKS the STATE.md write for step: 4-enforcement unless this marker exists with status: APPROVED — the same structural-gate-artifact pattern this skill mandates for created workflows, applied to itself.
If reviewer reports gaps → drive convergence via /goal Design reviewer returns APPROVED on .planning/wc/{name}/DESIGN.md. Stop after 5 turns. Each turn: fix DESIGN.md, re-dispatch the reviewer, end turn

Proceed to Step 4. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 4: Apply Enforcement Patterns

Context check: Steps 4-6 generate enforcement content and workflow files — the most context-intensive work. Before proceeding:

If context is low (≤35% remaining), write .planning/wc/{name}/HANDOFF.md with interview answers (from .planning/wc/{name}/INTERVIEW.md), phase decomposition (from .planning/wc/{name}/DESIGN.md), and current progress. Pause.
If context is critical (≤25% remaining), write HANDOFF.md immediately — do not start enforcement generation.

!cat ${CLAUDE_SKILL_DIR}/../../references/enforcement-checklist.md You MUST read this file before proceeding. No claiming you "remember" the patterns.

Step 4 is high-drift: you decide how much enforcement each generated phase carries, and the cheapest shortcut is to under-enforce.

| Excuse (Step 4) | Reality | Do Instead | |--------|---------|------------| | "This medium-drift phase doesn't really need an Iron Law" | Drift risk is a spectrum; the phase you under-enforce is the one that fails in production. | Assign enforcement by the tier table below, then write the actual content. | | "Listing the pattern names is enough" | The gate fails at Level 4: naming ≠ generating. A phase with "Iron Laws: yes" but no Iron Law text has none. | Draft the actual Iron Law / table / Red Flags per phase, not a checklist of names. | | "I'll make the mechanical rule a prompt line, faster than a hook" | Prompt rules drift and cost context; mechanical rules belong in hooks/.py. | Hook-or-.py for anything checkable; prose only for judgment. |

Drive check: under-enforcing here to "move faster to generation" is anti-helpful (the user inherits a workflow that drifts), incompetent (you skipped the drift-risk scoring that beats intuition), and anti-efficient (a missing gate costs 10× to fix during implementation). Enforce proportional to drift — that IS the deliverable.

For each phase, score which of the 13 patterns are needed:

High-drift phases (implementation, verification): Iron Laws, Rationalization Tables, Gate Functions, Drive-Aligned Framing, Artifact Review Gates
Medium-drift phases (design, review): Gate Functions, Red Flags, Staged Review Loops, Artifact Review Gates
Low-drift phases (brainstorm, exploration): Red Flags only (creative phases need freedom)

Generate the specific enforcement content:

Write Iron Laws with <EXTREMELY-IMPORTANT> tags
Build Rationalization Tables from the failure modes identified in Step 2
Define Red Flags + STOP for each phase's common wrong-path indicators

Hooks Over Prompt Enforcement

Before writing prompt-based enforcement for a constraint, ask: is this mechanically checkable? If yes, write a scoped hook instead.

Skills and agents can declare PreToolUse and PostToolUse hooks in their frontmatter. These hooks fire on every matching tool call during the skill's lifetime — no prompt tokens consumed, no drift, no rationalization.

For each constraint identified in the enforcement plan:

| If the constraint is... | Then use... | |------------------------|-------------| | Phase gate prerequisite | PreToolUse hook using phase-gate-guard.py — checks artifact exists with correct status before allowing Write/Edit/Agent | | File extension/path guard | PreToolUse hook on Read/Edit/Write — check path | | Tool parameter validation | PreToolUse hook — check required params | | Tool sequence enforcement | PreToolUse hook with state file — track what's been done | | Post-subagent restriction | PostToolUse on Agent + PreToolUse on restricted tools | | Quality/judgment call | Prompt enforcement (Iron Law, Red Flags) | | Educational/motivational | Prompt enforcement (Rationalization Table, Drive-Aligned Framing) |

Write the hook as a Python script in skills/[phase]/scripts/ and reference it in the skill's frontmatter:

hooks:
  PreToolUse:
    - matcher: "Read"
      hooks:
        - type: command
          command: "uv run python3 ${CLAUDE_PLUGIN_ROOT}/skills/[phase]/scripts/guard-media-files.py"

Design rule: Hook first. If the hook can't express the constraint (requires judgment, context, or semantics), fall back to prompt enforcement.

Deviation Rules for Implementation Phases

Any phase where agents execute work (implementation, drafting, transformation) should include a 4-rule deviation system governing unplanned discoveries:

| Rule | Trigger | Action | Permission | |------|---------|--------|------------| | 1: Bug | Broken behavior, errors, type errors, security vulns | Fix → test → verify → track [Rule 1 - Bug] | Auto | | 2: Missing Critical | Missing essentials: error handling, validation, auth, logging | Add → test → verify → track [Rule 2 - Missing Critical] | Auto | | 3: Blocking | Prevents completion: missing deps, wrong types, broken imports | Fix blocker → verify proceeds → track [Rule 3 - Blocking] | Auto | | 4: Architectural | Structural change: new service, schema change, switching libs | STOP → present decision → track [Rule 4 - Architectural] | Ask user |

Priority: Rule 4 (STOP) > Rules 1-3 (auto) > unsure → Rule 4

Adapt categories to the domain: For DS workflows, R1 includes data integrity bugs; R2 includes missing null handling; R4 includes schema changes. For writing workflows, R1 includes factual errors; R2 includes missing citations; R4 includes structural reorganization.

Each task summary should end with: Total deviations: N auto-fixed (R1: X, R2: Y, R3: Z). Impact: [assessment].

Gate: Enforcement Patterns Loaded [checkpoint: human-verify, auto-advanceable]

Verify enforcement-checklist.md was read
Check that you can name all 13 patterns
If you cannot list them, re-read enforcement-checklist.md

Update .planning/wc/{name}/STATE.md:

step: 4-enforcement
status: completed
implements: [WC-05]
requires: [DESIGN.md, enforcement-checklist.md]
provides: [enforcement pattern assignments per phase]
affects: [.planning/wc/{name}/STATE.md]
one-liner: "13 enforcement patterns assigned to phases based on drift risk."

Proceed to Step 4b. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 4b: Common Enforcement Across Skill Families

Context check: Cross-skill consistency analysis reads multiple sibling skills and produces hook/script coverage matrices. Before proceeding:

If context is low (≤35% remaining), write .planning/wc/{name}/HANDOFF.md with enforcement plan from Step 4 and current progress. Pause.
If context is critical (≤25% remaining), write HANDOFF.md immediately.

When multiple skills operate on the same domain, they need consistent enforcement across three layers: constraints (prompt), hooks (structural), and script wiring (gate orchestration). Scan the target plugin:

Layer 1: Shared Constraints (Co-located in `constraints/`)

List all skills/*/SKILL.md files in the target plugin directory
For each sibling skill, identify enforcement patterns (Iron Laws, Rationalization Tables, Red Flags)
Check if a references/constraints/ directory already exists with co-located .md + .py pairs

If constraints/ directory exists: new skills MUST Read() the specific .md files they need from that directory.

If no shared directory exists but sibling skills share the same domain: create references/constraints/ and extract common enforcement into atomic files — one .md per rule, co-located .py for mechanically testable rules. Each skill Read()s only the specific files it needs; skill-specific enforcement stays inline.

Constraint Propagation Rule: When adding a new rule, create the .md file (+ .py if testable) in references/constraints/ with applies-to frontmatter. Over-inclusion beats drift. The filesystem is the index — no separate TOC file to maintain.

Constraint Architecture: Co-located Pairs, Auto-discovered

Constraints and conventions are unit tests for agent behavior. The architecture follows test framework design: co-located files, auto-discovery, structured output, no manual wiring.

The One Directory

All rules live in a single constraints/ directory. No separate conventions/ directory. The distinction between constraint and convention is presence of a check script, not directory location.

references/
├── constraints/                       → all rules live here
│   ├── no-agent-resume.md             → rule (loaded into LLM context)
│   ├── no-agent-resume.py             → check script (run by test runner)
│   ├── source-first-fixes.md          → has .py pair = constraint (tested)
│   ├── source-first-fixes.py
│   ├── verbatim-quotes.md
│   ├── verbatim-quotes.py
│   ├── diagram-storytelling.md        → no .py pair = convention (judgment-only)
│   ├── section-transitions.md         → convention (graduation candidate)
│   └── ...

Same name links them. source-first-fixes.md + source-first-fixes.py = a constraint. diagram-storytelling.md alone = a convention. No frontmatter wiring. No manual registration. No index files to maintain.

Graduation = writing the .py file. A convention becomes a constraint the moment you add its check script. No moving files, no updating indexes.

The classification test: Ask the user: "Can you write a script that returns pass/fail for this rule?" If yes → write the .py file (constraint). If it requires reading and judging → .md only (convention). Some rules start as conventions and graduate when testability improves.

Examples:

no-agent-resume.md + no-agent-resume.py → constraint (mechanically detectable)
diagram-storytelling.md (no .py) → convention (requires judgment)
verbatim-quotes.md + verbatim-quotes.py → constraint (diff-checkable)
section-transitions.md (no .py) → convention (requires reading)

Rule File Structure

Each .md file is self-contained:

---
name: constraint-name
description: One-line trigger description
applies-to: [skill-1, skill-2, skill-3]
---

## Rule

The rule stated clearly.

## Rationale

**Why this exists** — cite the real incident or failure mode.

## Examples

### Correct
[Example of correct behavior]

### Incorrect
[Example of incorrect behavior]

## Rationalization Table

| Excuse | Reality | Do Instead |
|--------|---------|------------|
| ... | ... | ... |

## Red Flags

- **"..."** → STOP. [Why this thought is wrong]

Check Script Interface

Each .py file follows a standard interface so the runner can auto-discover and execute it:

#!/usr/bin/env python3
"""Constraint: no-agent-resume — NEVER use agent resume; spawn fresh."""

CONSTRAINT = "no-agent-resume"
APPLIES_TO = ["all"]
SEVERITY = "hard"  # hard = block, soft = warn

def check(context):
    """Returns list of violations. Empty list = pass."""
    violations = []
    # ... check logic ...
    return violations

if __name__ == "__main__":
    import json, sys
    violations = check({"cwd": sys.argv[1] if len(sys.argv) > 1 else "."})
    if violations:
        for v in violations:
            print(f"FAIL: {v}")
        sys.exit(1)
    print(f"PASS: {CONSTRAINT}")

Test Runner: Auto-discovery

The runner discovers check scripts — no manual wiring. Add a .py file, it runs automatically.

#!/usr/bin/env python3
"""check-all.py — auto-discovers and runs all constraint checks."""
import glob, importlib.util, json, sys

constraints_dir = "references/constraints"
results = {"passed": [], "failed": [], "conventions": [], "errors": []}

md_files = set(p.stem for p in Path(constraints_dir).glob("*.md"))
py_files = set(p.stem for p in Path(constraints_dir).glob("*.py"))

for name in sorted(md_files):
    if name in py_files:
        # Constraint — has check script, run it
        mod = import_check(f"{constraints_dir}/{name}.py")
        violations = mod.check(context)
        if violations:
            results["failed"].append({"name": name, "violations": violations})
        else:
            results["passed"].append(name)
    else:
        # Convention — no check script, flag for reviewer
        results["conventions"].append(name)

print(json.dumps(results, indent=2))
print(f"\n{len(results['passed'])}/{len(md_files)} passed, "
      f"{len(results['failed'])} failed, "
      f"{len(results['conventions'])} conventions (judgment-only)")
sys.exit(1 if results["failed"] else 0)

Coverage is automatic. The runner computes it from the filesystem — no hand-maintained coverage matrix.

Verification Architecture: Two Legs

Verification Phase
    ↓
Leg 1: uv run python3 check-all.py (auto-discovers constraints/*.py)
    ↓
    Structured results: {passed: [...], failed: [...], conventions: [...]}
    ↓                              ↓
    All passed                    FAIL → fix → re-run
    ↓
Leg 2: Spawn reviewer subagent
    ↓  (runner passes conventions list — the .md files without .py pairs)
    ↓  (reviewer loads those .md files and scores work against them)
    ↓                              ↓
    Score >= threshold             Score < threshold → revise → re-score
    ↓
VERIFIED

Both legs are necessary. Passing constraints but failing conventions = code that passes CI but fails code review. Passing conventions but failing constraints = code that looks good but has bugs.

How skills reference constraint files:

Skills use a bang to auto-load all applicable constraints at skill load time:

# In a skill's SKILL.md:
!`uv run python3 ${CLAUDE_SKILL_DIR}/../../scripts/load-constraints.py skill-name`

This mirrors check-all.py's auto-discovery but for .md prose:

#!/usr/bin/env python3
"""load-constraints.py — load .md constraint prose for a skill, filtered by applies-to."""
import yaml, sys
from pathlib import Path

def load(skill_name, constraints_dir="references/constraints"):
    for md in sorted(Path(constraints_dir).glob("*.md")):
        text = md.read_text()
        if not text.startswith("---"):
            continue
        _, fm, body = text.split("---", 2)
        meta = yaml.safe_load(fm)
        applies = meta.get("applies-to", [])
        if "all" in applies or skill_name in applies:
            print(f"\n{'='*60}")
            print(f"# Constraint: {meta.get('name', md.stem)}")
            print(f"{'='*60}")
            print(body.strip())

if __name__ == "__main__":
    load(sys.argv[1])

The script:

Globs constraints/*.md
Parses applies-to frontmatter
Filters for the skill name (or all)
Outputs concatenated content

Adding a new constraint = create the .md file with applies-to. No skill edits needed.

Fallback: For plugins without the loader script, explicit Read() calls to specific .md files still work — but the auto-discovery pattern is preferred.

No index file needed. The filesystem IS the index. ls constraints/*.md shows all rules. ls constraints/*.py shows all tests.

Constraint Propagation Rule: When adding a new rule, create the .md file in constraints/. If it's mechanically testable, also create the .py file. Set applies-to in the .md frontmatter. Over-inclusion beats drift.

Layer 2: Hook Coverage (Structural Enforcement)

For each sibling skill, extract the hooks: block from YAML frontmatter
Produce a Hook Coverage Matrix (skills × hooks):

| Hook Script | skill-1 | skill-2 | skill-3 | skill-4 |
|-------------|---------|---------|---------|---------|
| guard-a.py  | ✅ Pre  | ✅ Pre  | ❌      | ✅ Pre  |
| guard-b.py  | ✅ Post | ✅ Post | ❌      | ✅ Post |
| monitor.py  | ✅ Post | ✅ Post | ❌      | ✅ Post |

Flag any hook present in some siblings but not others
Require justification for intentional gaps (e.g., "router delegates immediately — hooks fire in the routed-to skill")

Why hooks drift: Hooks are added when a failure mode is discovered in one skill. The fix adds the hook to that skill's frontmatter but doesn't propagate to siblings. Unlike constraints (which can drift subtly), missing hooks are silent — no error, no warning, the enforcement just doesn't fire.

Layer 3: Script Wiring (Auto-discovery + Hooks)

With the co-located architecture, script wiring is simpler — the auto-discovering runner (check-all.py) handles batch execution. But hooks still need explicit wiring:

List all .py check scripts in references/constraints/
Verify the auto-discovering runner (check-all.py) exists and globs constraints/*.py
For guard hooks (PreToolUse/PostToolUse), verify each is in at least one skill's YAML frontmatter
Produce a Script Wiring Matrix:

| Script              | Auto-discovered | Hook Reference |
|---------------------|----------------|----------------|
| no-agent-resume.py  | ✅ constraints/ | ✅ pre-tool guard |
| source-first.py     | ✅ constraints/ | ✅ pre-tool guard |
| check-widows.py     | ✅ constraints/ | ✅ post-compile guard |
| new-check.py        | ✅ constraints/ | ❌ No hook |

Flag any .py file in constraints/ without a corresponding .md file (orphaned test)
Flag any guard hook that references a script NOT in constraints/ (manual wiring when it should be co-located)

Why auto-discovery eliminates most wiring bugs: Adding a .py to constraints/ = automatically run by the test runner. No manual registration. The main wiring failure mode now is hooks — guard hooks still need explicit YAML frontmatter.

Gate: Cross-Skill Consistency Complete [checkpoint: human-verify, auto-advanceable]

Verify sibling skills were scanned (or note that no siblings exist)
Layer 1: If constraints/ directory exists, verify sibling skills Read() the specific .md files they need. If skills share a domain, verify common rules are in constraints/ (not inlined).
Layer 2: Hook Coverage Matrix produced. No unexplained gaps.
Layer 3: Script Wiring Matrix produced. No unwired scripts.

Update .planning/wc/{name}/STATE.md:

step: 4b-cross-skill
status: completed
implements: [WC-05]
requires: [DESIGN.md, enforcement-checklist.md]
provides: [enforcement plan, hook coverage matrix]
affects: [.planning/wc/{name}/STATE.md]
one-liner: "Hook coverage and script wiring matrices produced across the skill family; intentional gaps justified."

Proceed to Step 5. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 5: Design Two Entry Points

Context check: Entry point design is moderate effort. Before proceeding:

If context is low (≤35% remaining), write .planning/wc/{name}/HANDOFF.md with enforcement plan and current progress. Pause.
If context is critical (≤25% remaining), write HANDOFF.md immediately.

Red Flags — STOP: designing only one entry point (the fresh-start one) and skipping the midpoint · writing a midpoint that loads a summary of constraints instead of Read()-ing the actual .md files (summaries enable reward-hacking — the agent checks a 4-item digest, finds nothing, reports "all clear") · a midpoint that depends on prior-phase context instead of being self-contained.

Every workflow exposes exactly two user-facing commands. Everything else is internal.

| Entry Point | Purpose | Example | |-------------|---------|---------| | Entry (start fresh) | Begins a new episode, runs brainstorm phase first | /dev, /ds, /writing | | Midpoint (re-enter) | Re-enters a running episode, diagnoses and routes to the right phase | /dev-debug, /ds-fix, /writing-revise |

Why two: The user never needs to know which internal phase to invoke. Entry starts fresh. Midpoint diagnoses what's wrong and routes.

Midpoint Constraint Loading

The entry point runs sequentially — each phase loads its constraints and passes context forward. The midpoint can't rely on that. It may run in a new session, after context compression, or hours after the last edit. Prior constraints are gone.

The midpoint must be self-contained. It loads every constraint layer it needs before touching the work:

/writing-revise loads:
  1. .planning/ACTIVE_WORKFLOW.md    → workflow state (what phase, what style)
  2. .planning/PRECIS.md, .planning/OUTLINE.md → structural intent (what we're building)
  3. ai-anti-patterns      → universal constraints (no AI-smell)
  4. domain skill           → domain constraints
  THEN: check the draft against all four layers

/dev-debug loads:
  1. .planning/HYPOTHESES.md          → what's been tried
  2. .planning/LEARNINGS.md           → accumulated knowledge
  THEN: spawn fresh subagent for next investigation iteration

/ds-fix loads:
  1. .planning/SPEC.md, .planning/PLAN.md       → objectives and task breakdown
  2. .planning/LEARNINGS.md            → pipeline state and observations
  3. output-first protocol   → verification enforcement
  THEN: diagnose and route to fix path

Critical rule: Any phase that evaluates quality must load the full constraint set, not a summary of it. Summaries enable reward hacking — the agent checks against a 4-item summary, finds no issues, and reports "all checks pass" when the full rules would have caught problems. The fix: Read() the actual skill before checking.

Shared Constraint Files

See Step 4b for the full atomic constraint/convention architecture. This section covers only the midpoint-specific concern.

Midpoint constraint loading: The midpoint must load every constraint it needs before touching the work. Load the specific .md files relevant to the current phase directly — no index file needed.

/writing-revise loads:
  1. .planning/ACTIVE_WORKFLOW.md    → workflow state
  2. references/constraints/verbatim-quotes.md → specific constraint for revision
  3. references/constraints/source-first-fixes.md → specific constraint for revision
  THEN: check the draft against loaded constraints

Session Handoff Support

Both entry points should support session handoff via .planning/HANDOFF.md — a structured pause/resume mechanism for when work spans multiple sessions.

Entry point startup check:

1. Check if .planning/HANDOFF.md exists
2. If found → read it, offer to resume from recorded state
3. If not found → proceed with normal entry (fresh start or midpoint diagnosis)

Handoff document requirements:

YAML frontmatter (phase, task, status, last_updated) for machine parsing
Sections: Current State, Completed Work, Remaining Work, Decisions Made, Rejected Approaches, Blockers, Next Action
"Next Action" must be specific enough to start immediately (not "continue working")

Why: Long workflows often exceed context windows. Without structured handoff, the next session wastes significant time re-discovering where the previous session left off. The handoff captures decisions, dead ends, and in-flight context that state files alone don't preserve.

Gate: Two Entry Points Designed [checkpoint: human-verify, auto-advanceable]

Verify entry point (start fresh) is defined
Verify midpoint (re-enter) is defined with constraint loading
If either is missing, design both entry points

Update .planning/wc/{name}/STATE.md:

step: 5-entry-points
status: completed
implements: [WC-06]
requires: [DESIGN.md]
provides: [entry point design, midpoint design]
affects: [.planning/wc/{name}/STATE.md]
one-liner: "Entry point (start fresh) and midpoint (re-enter with constraint loading) designed."

Proceed to Step 6. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 6: Generate Workflow Files

<EXTREMELY-IMPORTANT> **NO GENERATED PHASE FILE WITHOUT THE ENFORCEMENT DENSITY ITS DRIFT TIER DEMANDS.** Step 6 is the highest-drift phase of Mode 1 — you are now WRITING, and the temptation is to emit a clean-looking SKILL.md that silently drops the gate, the Iron Law, or the co-located `.py` that DESIGN assigned. A generated file that looks complete but omits its enforcement ships the gap to every user of that workflow. </EXTREMELY-IMPORTANT>

| Excuse (Step 6) | Reality | Do Instead | |--------|---------|------------| | "The design looks complete, I'll skip enforcement on the medium-drift phase" | Medium-drift phases still drift. The omission is invisible until the workflow fails. | Generate the enforcement DESIGN assigned to EVERY phase, by tier. | | "I'll write the constraint .md now and the .py later" | Later never comes; the runner auto-discovers nothing. | Write the .md + .py together — same stem, same dir. | | "Transitions as prose are fine" | Prose transitions are advisory; users invoke phases directly and bypass them. | Wire hook-enforced gate artifacts where DESIGN marked them mandatory. |

Red Flags — STOP if you catch yourself: about to write a phase SKILL.md without the gate DESIGN specified · about to skip the co-located .py for a testable rule · about to write a verifier/reviewer agent without read-only allowed-tools · about to emit a hook command: with ${CLAUDE_SKILL_DIR} instead of ${CLAUDE_PLUGIN_ROOT}. Drive check: skipping enforcement to "ship the files faster" is anti-helpful — the user inherits every gap you dropped.

Create the following artifacts:

Entry command (skills/[name]/SKILL.md) — routes to first phase
Midpoint command (skills/[name]-fix/SKILL.md or skills/[name]-debug/SKILL.md) — self-contained re-entry
Phase skills (skills/[name]-[phase]/SKILL.md) — one per phase, internal only
Constraint files — co-located in references/constraints/:
- One .md per rule (loaded into LLM context)
- Co-located .py for mechanically testable rules (auto-discovered by runner)
- .md without .py = convention (judgment-only)
- check-all.py auto-discovering runner
Wire up transitions — each phase ends by reading the next phase's skill
Post-subagent boundaries — each phase skill that dispatches a subagent MUST include:
- What main chat CAN do after subagent returns (read results, fix gaps, present to user)
- What main chat CANNOT do (re-score, override subagent findings, declare "close enough")
- Tool restriction tier for the subagent (executor/verifier/researcher/auditor)

State Folder Convention

Workflows should store all state files in a .planning/ directory at the project root (not .claude/). This keeps workflow state separate from Claude Code configuration.

Standard state files (all written to .planning/): | File | Purpose | When Created | |------|---------|-------------| | .planning/SPEC.md | Requirements, goals, constraints | Brainstorm/clarify phase | | .planning/PLAN.md | Task breakdown with status tracking | Design phase | | .planning/STATE.md | Current workflow position (active phase, blockers) | Entry point startup | | .planning/HANDOFF.md | Session pause/resume context | On pause or context exhaustion | | .planning/VALIDATION.md | Requirement-to-test coverage map | Validation phase | | .planning/LEARNINGS.md | Accumulated discoveries and decisions | Throughout workflow |

Design principles: File-based, git-trackable, human-editable. No databases, no external services. YAML frontmatter for machine-readable state; markdown body for human reading.

Visual Output for Human Verification (Learn-by-Doing)

Visual artifacts can make decision checkpoints faster — but what helps depends on the human and the domain. Don't prescribe visual output. Observe what the human actually does during review, then offer to automate it.

The learning pattern:

Observe — at each decision checkpoint, note what the human asks for. Do they want a diff? A table? A chart? Do they open files in another tool? Or do they just read the summary and approve?
Record — log the observation in LEARNINGS.md: "User evaluated results by asking for coefficient comparison table" or "User approved after reading test summary"
Offer — after 3+ reviews with the same pattern, offer to bundle a script that generates the view automatically. Don't build it speculatively.

When visual output IS worth building:

The human has asked for the same view 3+ times
The checkpoint involves evaluating a distribution or pattern (spec curves, coverage maps) — humans do visual pattern recognition faster than reading tables
The output is a rendered artifact (slides, documents) where "does it look right?" is the literal question

When visual output is NOT worth building:

The human reads the summary and approves — that's fine, don't add friction
The checkpoint is a yes/no with clear criteria (tests pass, file exists)
Building the visualization takes longer than the verification it replaces

Available patterns (offer when the human's review behavior suggests them):

| If the Human Asks For... | Consider Building | |--------------------------|-------------------| | "Show me what changed" | Interactive diff explorer (HTML) | | "What's the architecture?" | Dependency graph / codebase tree | | "Are results robust?" | Specification curve (specr/marimo) | | "Where did we lose rows?" | Row count waterfall chart | | "What changed since last draft?" | DOCX redline / tracked changes | | "Does the draft cover all claims?" | Claim coverage heatmap |

Implementation: bundle scripts in skills/[phase]/scripts/. Self-contained HTML or notebooks. The verify/review phase offers to run the script — it doesn't force it.

Present complete file list for user approval before writing. [checkpoint: decision — user chooses which files to generate]

Optional: generate the files with the wc-generate transform workflow

Once the user has approved the file list, the per-file creation from the approved DESIGN.md is a textbook transform/generate fan-out — a fixed spec (DESIGN.md) drives one independent write per file, with no creative latitude. workflow-creator eats its own cooking here too: prefer the wc-generate ultracode workflow over hand-writing each file in main chat. (The interview, decomposition, enforcement design, and the user file-approval gate above stay CONVERSATIONAL — only the mechanical per-file write moves into the workflow.)

WF=$(command ls -d ~/.claude/plugins/cache/edwinhu-plugins/workflows/*/workflows/wc-generate.js 2>/dev/null | sort -V | tail -1)
[ -z "$WF" ] && WF="${CLAUDE_SKILL_DIR}/../../workflows/wc-generate.js"

Workflow({ scriptPath: "<WF>", args: {
  workflowName: "{name}",
  projectDir: "<abs plugin repo root>",
  designPath: "<abs .planning/wc/{name}/DESIGN.md>"   // optional; defaults to that path
} })

It fans out one worktree-isolated write-agent per file (each phase SKILL.md + each constraint .md/.py pair + check-all.py), each creating its file from the pinned DESIGN spec (the workflow refuses if DESIGN.md is missing/unapproved — Delete & Restart). A read-only verify stage then confirms each file matches its spec and the co-located .md/.py pairing holds. It returns { overallPass, scoreTable, files, findings, filesThatFailed }.

Ground-truth before claiming done (self-reports are not ground truth): the write-agents run in isolated worktrees, so after the workflow returns you MUST (1) merge the surfaced worktrees, then (2) ls each file at its expected path (watch for doubled skills/skills/ or workflows/workflows/ typo paths) and node --check / lint the generated files yourself. On a re-run after fixing gaps, pass onlyChecks: <prev result.filesThatFailed> + priorReviews: <prev result.reviews>. If wc-generate is unavailable, fall back to writing the files directly per the list above. Whichever path you take, Step 7's wc-audit self-audit still runs on the result.

Update .planning/wc/{name}/STATE.md:

step: 6-generate
status: completed
implements: [WC-07, WC-10, WC-11]
requires: [INTERVIEW.md, DESIGN.md]
provides: [skills/{name}/SKILL.md, skills/{name}-fix/SKILL.md, phase skills, constraint files]
affects: [skills/{name}/, references/constraints/]
key-files: {created: [skills/{name}/SKILL.md, skills/{name}-fix/SKILL.md]}
one-liner: "All workflow files generated: entry command, midpoint, phase skills, constraints."

Deviation Rules for Steps 4-6 (File Generation)

During enforcement generation (Step 4) and file writing (Step 6), unplanned issues may arise. Apply these deviation rules:

| Rule | Trigger | Action | Permission | |------|---------|--------|------------| | R1: Bug | Broken path, invalid YAML, syntax error in generated file | Fix immediately, note in STATE.md | Auto | | R2: Missing Critical | Generated workflow missing a gate, missing enforcement for high-drift phase | Add the missing element, note in STATE.md | Auto | | R3: Blocking | Constraint file conflict, skill naming collision, hook script missing | Fix blocker, note in STATE.md | Auto | | R4: Architectural | User's domain doesn't fit proposed topology, need to restructure phases | STOP — present decision to user with options | Ask user |

Priority: R4 (STOP) > R1-R3 (auto) > unsure → R4

Delete & Restart Protocol

<EXTREMELY-IMPORTANT> If you generate workflow files (Step 6) without having completed Steps 1-5 — or if STATE.md does not show steps 1-5 as completed — DELETE all generated files and restart from Step 1. No exceptions.

Patching files generated without proper investigation, interview, decomposition, and enforcement design is worse than restarting. The generated files inherit every gap from the skipped steps. </EXTREMELY-IMPORTANT>

Step 7: Self-Audit the Generated Workflow

Traceability (self-applied P18): Before/with the self-audit, write .planning/wc/{name}/VALIDATION.md mapping each WC-NN requirement (from the  tags in the generated skills) to the concrete evidence that verifies it — the gate, hook, or audit check that confirms it — and tag each requirement's scope (v1 / v2 / out-of-scope). This closes the loop from requirement → verification evidence, not just requirement → spec.

Review-pattern logging (self-applied P19b): At each present-to-user checkpoint (Step 6 file presentation, Step 7 audit results), append to .planning/wc/{name}/LEARNINGS.md what the user attended to or changed (which findings they prioritized, what they overrode). This is the observe→record→offer loop wc prescribes for the workflows it creates, applied to wc itself — after 3+ recurring patterns, propose encoding them as defaults.

Context check: Step 7 dispatches a subagent with a large prompt containing all generated file paths and audit criteria. This is one of the most context-intensive operations. Before proceeding:

If context is low (≤35% remaining), write .planning/wc/{name}/HANDOFF.md with generated file list, current step, and note that self-audit is pending. Pause.
If context is critical (≤25% remaining), write HANDOFF.md immediately — do not attempt the subagent dispatch.

<EXTREMELY-IMPORTANT> ## The Iron Law of Eating Your Own Cooking

NO GENERATED WORKFLOW WITHOUT A MODE 2 AUDIT ON IT. This is not negotiable.

workflow-creator mandates audit-fix loops, independent verification, and artifact review gates for every workflow it creates. It cannot exempt its own output from these same standards.

Skipping the self-audit is NOT HELPFUL — you're shipping an unverified workflow to the user. The user will discover the gaps when the workflow fails in production. The 5-minute audit would have caught them. </EXTREMELY-IMPORTANT>

After generating workflow files in Step 6:

<EXTREMELY-IMPORTANT> **The audit MUST be run by the wc-audit workflow (read-only reviewers, JS gate) — not by you.** If you score the files yourself, you are self-reviewing your own work; see the Iron Law above. The same agent that wrote the files cannot score them independently. </EXTREMELY-IMPORTANT>

Run the wc-audit workflow on the newly generated workflow (same workflow Mode 2/Mode 3 use — read-only reviewers, JS-computed gate, so generation and judgment are structurally separate). First drafts clear 8.0 (composite), not the calibrated ceiling — but they still must reach substratePass (0 critical, no enforcement Absent, portability Clean) before presenting:
```
WF=$(command ls -d ~/.claude/plugins/cache/edwinhu-plugins/workflows/*/workflows/wc-audit.js 2>/dev/null | sort -V | tail -1)
[ -z "$WF" ] && WF="${CLAUDE_SKILL_DIR}/../../workflows/wc-audit.js"
```
```
Workflow({ scriptPath: "<WF>", args: { targetWorkflow: "{name}", projectDir: "<abs repo root>", pluginRoot: "<abs .../workflows dir>", threshold: 8.0 } })
```
Write result.reportMarkdown to .planning/wc/{name}/AUDIT.md and append result.composite to .planning/wc/{name}/SCORES.md.
Check score: If result.substratePass is false OR result.composite < 8.0, drive convergence via the native /goal primitive — a separate evaluator gates exit by reading SCORES.md from the transcript, so the agent that generated the files isn't also the judge.

Invoke:
```
/goal Generated workflow reaches result.substratePass=true (0 critical, no enforcement Absent, portability Clean) AND composite >= 8.0 in .planning/wc/{name}/SCORES.md from a fresh audit subagent. Stop after 3 turns.
```
Each turn under the active goal: fix the generated files based on the latest AUDIT.md findings, re-run the wc-audit workflow (full pass, or onlyChecks: <prev result.reviewersThatFlagged> + priorReviews: <prev result.reviews>), append the new composite to SCORES.md, end turn.
Present to user [checkpoint: decision — user approves or requests changes] with the audit report attached — the user sees both the workflow AND its quality score

Step 6: Generate Files
    ↓
Step 7: Mode 2 Audit on generated files
    ↓
Score >= 8.0? ──YES──→ Present files + audit report to user
    │
    NO
    ↓
/goal-driven fix loop (3-turn budget) → fresh audit subagent re-scores each turn
    │
    ↓ (after 3-turn budget elapses)
Present files + audit report + remaining gaps to user

Why an 8.0 composite floor (not the ceiling) at Step 7: Generated workflows are first drafts; the composite needs real-world iteration to approach the calibrated ceiling (~9.0). But a first draft must already be structurally sound — substratePass true (no missing gates, no broken paths, no ungated transitions) and composite ≥ 8.0. Mode 3 then drives the 8.0 → substrate-clean-at-ceiling climb (NOT a chase to 9.5 — see the Iron Law of Workflow Improvement).

Post-Subagent Enforcement (Step 7)

After the audit subagent returns, main chat operates under these restrictions:

| Main chat CAN do (verification) | Main chat CANNOT do (investigation) | |----------------------------------|--------------------------------------| | Read AUDIT.md and SCORES.md | Re-score principles (auditor's scores are authoritative) | | Fix specific gaps identified in AUDIT.md | Override audit findings ("the auditor was wrong about X") | | Re-dispatch a NEW audit subagent | Declare "close enough" below 8.0 | | Present results to user | Edit generated files without addressing a scored gap |

The audit subagent's score is authoritative. If you disagree with a score, fix the gap and let the next audit re-score — do not override.

Update .planning/wc/{name}/STATE.md:

step: 7-self-audit
status: completed
implements: [WC-08, WC-09]
requires: [DESIGN.md, generated skill files]
provides: [AUDIT.md]
affects: [.planning/wc/{name}/AUDIT.md]
key-files: {created: [.planning/wc/{name}/AUDIT.md]}
one-liner: "Fresh subagent audit complete. Composite score and per-principle gaps recorded."

Mode 2: Audit Existing Workflow

Step 1: Read All Files ──→ Step 2: Score 20 Principles ──→ Step 3: Score 13 Patterns
  [auto]                     [auto]                          [auto]
    │                           │                               │
    ▼                           ▼                               ▼
  File map built              P01-P21 scored                  Patterns scored per phase
                                                                  │
                              Step 3b: Path Portability ◄─────────┘
                                [auto]
                                  │
                                  ▼
                              Step 4: Output Report
                                [decision: present to user]
                                  │
                                  ▼
                              AUDIT.md written

<EXTREMELY-IMPORTANT> **NO PAUSE BETWEEN AUDIT STEPS.** Complete Step 1 → 2 → 3 → 3b → 4 without stopping. Do NOT pause or wait for user input between steps. Pausing mid-audit resets working memory, anchors scores on the last principle read, and produces scores that reflect reading order rather than the evidence — the exact failure Mode 2's Delete & Restart exists to undo. </EXTREMELY-IMPORTANT>

State initialization: Create .planning/wc/{name}/STATE.md with this YAML template (the wc-state-frontmatter constraint requires requires/provides/affects):

---
mode: audit
step: 1-read
status: in_progress
target: [workflow name]
implements: [WC-09]
requires: [all target workflow skill files]
provides: [file map]
affects: [.planning/wc/{name}/STATE.md]
one-liner: "Audit started on {target} — discovering skill files via wc-audit."
---

Context monitoring: Mode 2 audits complex multi-file workflows. Check context availability:

If context is low (≤35% remaining), write .planning/wc/{name}/HANDOFF.md and pause — the audit will degrade if context is exhausted mid-scoring.
If context is critical (≤25% remaining), write HANDOFF.md immediately.

<EXTREMELY-IMPORTANT> ### How Mode 2 runs: the wc-audit ultracode workflow (eat your own cooking)

The audit fan-out is owned by an ultracode workflow — a script, not hand-dispatched agents in main chat. workflow-creator tells every other workflow to migrate its review fan-out to a Claude Code ultracode workflow; it MUST do the same for its own audit. workflows/wc-audit.js fans out one read-only reviewer per audit dimension (4 architecture clusters covering P01-P21, the 13-pattern enforcement checklist, path portability, the Ultracode-Workflow Candidacy Scan), adversarially verifies each critical/major gap against the actual files, and computes the composite + verdict in pure JS — so the model can no longer self-report a generous composite (the exact honor-system smell this skill flags in others). Steps 1-3b below are the criteria the workflow's reviewers read (they Read this SKILL.md's Mode 2 section for the P01-P21 definitions); Step 4 is how you render AUDIT.md from the result — you do NOT score by hand.

Run it:

Resolve the cached workflow path (local-plugin fallback when running from source):

WF=$(command ls -d ~/.claude/plugins/cache/edwinhu-plugins/workflows/*/workflows/wc-audit.js 2>/dev/null | sort -V | tail -1)
[ -z "$WF" ] && WF="${CLAUDE_SKILL_DIR}/../../workflows/wc-audit.js"
echo "$WF"

Run it (full audit first; on a Mode 3 re-audit pass onlyChecks + priorReviews from the prior result):

Workflow({ scriptPath: "<WF>", args: {
  targetWorkflow: "{name}",
  projectDir: "<abs plugin repo root>",
  pluginRoot: "<abs .../workflows dir>"   // optional; helps resolve enforcement-checklist.md + the migration playbook
} })

It returns { overallPass, composite, verdict, threshold, isMetaTool, scoreTable, reportMarkdown, candidacyTable, findings, reviews, reviewersThatFlagged }. The reviewers ground in the criteria in Steps 1-3b; the JS gate computes the composite as the mean of non-exempt, non-domain-ceiling principle scores and honors the meta-tool exemptions (when auditing workflow-creator itself, P01 and P06 are excluded from the composite — see the structure note at the top of this skill).

Post-workflow boundary (verification, not investigation): after the workflow returns you may Read AUDIT.md/result.*, render the report, and fix gaps — you may NOT recompute or rationalize result.composite/result.overallPass (the JS owns the arithmetic), nor re-score a principle the reviewers scored.

The STATE.md step-chain (1-read → 2-score → 3-enforcement → 3b-portability → 4-report) is preserved — the workflow performs the work each step describes; you still write each STATE.md transition so the hook chain holds. This step-chain IS Mode 2's structural gate enforcement (P03): wc-step-gate-guard.py Layer 2 BLOCKS writing step: N to STATE.md unless step: N-1 shows status: completed — a tool-call-layer block, not advisory prose, and the reason a separate per-step marker file is unnecessary for Mode 2's linear chain (the same Layer-2 enforcement governs Mode 3's 1-initial-audit → 1-audit-loop). </EXTREMELY-IMPORTANT>

Step 1: Read the Workflow

The wc-audit workflow's Discover phase reads the target workflow's entry command, midpoint, and ALL phase skills (plus references/constraints) and returns the file map — you do not need to read them all into main context. Confirm the workflow's discovery enumerated the full file set (the workflow throws if it found none). Build/Read the phase map only as needed to interpret the result.

Gate: Workflow Fully Read [checkpoint: human-verify, auto-advanceable]

Verify entry command was read
Verify ALL phase skills were read (count Read() calls)
If any phase skill is missing, read it now

Update .planning/wc/{name}/STATE.md:

step: 1-read
status: completed
requires: [all workflow skill files]
provides: [file map, phase/transition inventory]
affects: [.planning/wc/{name}/STATE.md]
one-liner: "wc-audit Discover enumerated the target's entry/midpoint/phase skills + references — full file map built."

Proceed to Step 2. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 2: Score Against Core Principles (P01-P21)

<EXTREMELY-IMPORTANT> **Scoring is the highest-drift step in this skill.** You are about to score 22 principles across multiple files and the pull is to skim, anchor to a first impression, and award generous round numbers. A generous score here is the most damaging error wc makes: it ships an improvement plan built on a false baseline. </EXTREMELY-IMPORTANT>

| Excuse (while scoring) | Reality | Do Instead | |--------|---------|------------| | "This principle is obviously a 9, I can tell from the structure" | "Can tell from structure" is the anchoring that produced a generous 6.5 where the careful tally was 5.2 (Apr 2026). | Cite the specific file:line that earns the score. No line → the score is a guess. | | "It's mostly there, call it 8" | "Mostly" means a gap exists; an honest score reflects the gap, not the vibe. | Find the gap, score to it, write the one-line justification. | | "I'll reuse my impression from the last principle" | Adjacent-principle anchoring makes scores reflect reading order, not evidence. | Score each principle from its own evidence; reset between principles. |

Red Flags — STOP: awarding a 9-10 without a cited line · scoring before reading the relevant file section · rounding a 7-ish up to 8 to "move on" · trusting a self-reported composite (the JS gate owns it). Drive check: a fast generous audit is anti-helpful — the user acts on a false baseline and the workflow fails where you said it was fine.

Score each principle 0-10. Use the formal ID (P01-P21) in all audit output for traceability.

P01 — Phased decomposition:

Does each phase have a single responsibility?
Are phase boundaries clear?
Can phases be executed out of order? (they shouldn't be)

P02 — Gates (deterministic or judgment-based):

Are gates machine-verifiable where possible? (file exists, test passes)
For subjective domains, are judgment gates explicit? (agent-assessed or human-assessed)
Or are they just prose? ("ensure quality is high")
Are there ungated transitions?

P03 — Structural gate enforcement (CRITICAL — this is the #1 audit gap):

For every mandatory inter-phase gate, classify as STRUCTURAL or ADVISORY:
- STRUCTURAL: Producing phase writes a concrete artifact (.planning/X_REVIEWED.md), consuming phase checks for it at startup and refuses to proceed without it
- ADVISORY: Gate uses instructional text only ("you must", "prerequisite:", "do not proceed without") — no artifact, no check
Any advisory-only mandatory gate is a defect. Flag it in the Critical Gaps section.
Check BOTH sides: (1) does the producing phase actually write the artifact? (2) does the consuming phase actually check for it?
Produce a Gate Enforcement Matrix:

| Transition | Gate | Artifact | Producer Writes? | Consumer Checks? | Status |
|------------|------|----------|-------------------|-------------------|--------|
| design → implement | plan reviewed | PLAN_REVIEWED.md | ✅ | ✅ | STRUCTURAL |
| explore → clarify | exploration done | (none) | ❌ | ❌ | ADVISORY ⚠️ |

Additionally classify STRUCTURAL gates as HOOK-ENFORCED or INSTRUCTION-ONLY:
- HOOK-ENFORCED: Skill frontmatter declares a PreToolUse hook that checks for the artifact (strongest)
- INSTRUCTION-ONLY: Skill text checks for the artifact but no hook blocks tool calls (weaker — can be rationalized past under context pressure)
Score: count of STRUCTURAL gates / total mandatory gates. Below 80% = critical gap. Count of HOOK-ENFORCED / STRUCTURAL gates — below 50% = recommend hook migration.

P04 — Independent verification:

Is verification structurally independent from implementation? (fresh subagent, not self-review)
Does the verifier see only spec + output, not the implementation journey?
For subjective output, are there multiple specialized reviewers? (team topology)
Is self-review ever the final gate? (it shouldn't be)
Does verification check all 4 depth levels, or just existence?
Does any agent spawn its own sub-agents? (nested dispatch — must be flat instead)

Verification depth levels (from GSD goal-backward verification):

| Level | Name | Checks | Example Failure | |-------|------|--------|-----------------| | 1 | Exists | File/function/test physically present | Test file never created | | 2 | Substantive | Not a stub, placeholder, or TODO | Function body is pass or return {} | | 3 | Wired | Connected to the system (imported, called, routed) | Component defined but never rendered | | 4 | Functional | Actually works end-to-end | Tests pass, feature runs |

If verification only checks Level 1 (exists), it's theater. A workflow that claims "test exists" without checking the test is substantive, wired, and functional is shipping false confidence.

P05 — Artifact review:

Are intermediate artifacts (specs, plans, outlines) reviewed before downstream phases consume them?
Is the reviewer a fresh subagent (not the phase that wrote the artifact)?
Is there a fix-and-re-review loop with iteration limits?
Are large artifacts (>15 items) chunked for separate review?
Is there model tier guidance for delegation phases?

P06 — Two entry points:

Does the workflow have both an entry (start fresh) and midpoint (re-enter)?
Is the midpoint self-contained? (loads all constraints, doesn't depend on prior phases)
Does the midpoint load full skills, not summaries?
Do skills that share a domain share a common enforcement file? (or does each skill enforce its own version of the rules?)
Could a user get inconsistent enforcement depending on which skill they invoke?

P07 — Cross-skill consistency (three layers):

Constraints: Do all sibling skills Read() from the same constraints/ directory? Are rules co-located (.md + .py pairs for testable rules, .md only for conventions)? Is there an auto-discovering runner (check-all.py) that globs constraints/*.py?
Hooks: Do all sibling skills declare the same hooks in their YAML frontmatter? If a hook is present in some siblings but not others, is the gap justified? (Produce a Hook Coverage Matrix: skills × hooks)
Script wiring: Is every check script referenced in all three layers: (a) hook frontmatter, (b) batch orchestrator, (c) verification-checks definition? (Produce a Script Wiring Matrix: scripts × invocation points)

P08 — Constraint/convention test coverage:

Do all rules live in a single constraints/ directory? (no separate conventions/ directory)
Is the constraint/convention distinction based on presence of a .py check script, not directory location?
Does check-all.py (auto-discovering test runner) exist? Does it glob constraints/*.py — no manual wiring?
Does the verification phase run both legs: constraint checks (test runner, hard block) AND convention scoring (reviewer subagent loads .md files without .py pairs, soft block)?
Are there .md-only files (conventions) that could graduate to constraints by adding a .py check script?
Compute coverage from the filesystem: len(*.py) / len(*.md) — what percentage of rules have mechanical tests?

P09 — Iteration strategy:

Does each phase have an appropriate iteration topology? (one-shot, serial, parallel, team)
Are exit conditions structural (tests, convergence, human approval) not honor-system (promises)?

P10 — Post-subagent enforcement (from dev-debug v5.0 audit, March 16 2026):

When a subagent returns, what is main chat allowed to do? Is there an explicit tool whitelist?
Is "verification" defined concretely for this domain? (Without a definition, investigation gets disguised as verification)
Are operational tools (Bash commands beyond test running, Read on source files, Grep/Glob) restricted after subagent returns?
Is there a topic change protocol? (Without one, off-topic user messages silently kill iterative loops)

| Domain | Verification (main chat CAN do) | Investigation (main chat CANNOT do) | |--------|----------------------------------|--------------------------------------| | Dev | Run test suite, check exit code | Read source, grep, docker exec, curl, log reading | | DS | Check output file exists, view summary stats | Re-run queries, explore data, read notebook cells | | Writing | Read review summary artifact | Read/edit the draft, rephrase sections, "polish" |

The post-subagent moment is the highest-risk point in any delegated workflow. If the audit finds no enforcement there, flag it as a critical gap.

P11 — Deviation rules (from GSD 4-rule system):

Do implementation phases have a deviation rule system (auto-fix for bugs/missing/blocking, STOP for architectural)?
Are deviation categories adapted to the domain?
Are deviations tracked and summarized per task?

P12 — State management:

Does the workflow use .planning/ for state files (not .claude/ or scattered locations)?
Are standard state files present (.planning/SPEC.md, .planning/PLAN.md, .planning/STATE.md, .planning/LEARNINGS.md)?
Is state file-based, git-trackable, and human-editable?

P13 — Session handoff:

Does the entry point check for .planning/HANDOFF.md on startup?
Is the handoff document structured with frontmatter and mandatory sections?
Can work resume from a handoff without re-discovering context?

P14 — Checkpoint types:

Are gates classified by type (human-verify, decision, human-action)?
Can the workflow auto-advance human-verify checkpoints in autonomous mode?
Are true decision points (multiple valid approaches) distinguished from rubber-stamp approvals?

P15 — Context monitoring:

Do phases check context availability before starting expensive work?
Is there a handoff trigger when context is low (≤35%)?
Does the workflow degrade gracefully or just produce garbage at context exhaustion?

P16 — Summary frontmatter:

Do phase completions produce structured YAML summaries?
Do summaries include implements, requires, provides, affects fields?
Is the one-liner substantive (not "Phase complete")?

P17 — Agent tool restrictions:

Are verification/review agents restricted to read-only tools via allowed-tools frontmatter?
Can a verifier Write or Edit? (it shouldn't — that bypasses plan-execute-verify)
Are tool restriction tiers appropriate for each agent role?

P18 — Requirement traceability:

Do requirements have unique IDs in .planning/SPEC.md (e.g., AUTH-01)?
Do .planning/PLAN.md tasks reference requirement IDs?
Does .planning/VALIDATION.md map every ID to test evidence?
Is there a scope classification (v1/v2/out-of-scope)?

P19 — Autonomous phase chaining:

Can phases chain automatically without human intervention at every step?
Does the workflow batch ambiguities (smart discuss) instead of sequential asks?
Does it re-read the plan after each phase to catch dynamically inserted phases?
Are blockers handled with retry/skip/stop options?

P19b — Visual output for human verification:

Do decision checkpoints offer visual artifacts when the human's review pattern suggests them?
Does the workflow log what the human actually looks at during review (in .planning/LEARNINGS.md)?
If the human has asked for the same view 3+ times, has it been automated into a script?

P20 — Hooks over prompt enforcement:

Are mechanically-checkable constraints enforced via scoped hooks (PreToolUse/PostToolUse in skill frontmatter)?
Or are they enforced only via prompt text (Iron Laws, Red Flags) that consume context and can be rationalized away?
Specifically check for: phase gate enforcement (prerequisite artifact checks), file extension guards, path guards, tool parameter validation, tool sequence enforcement, post-subagent restrictions
Behavioral/motivational constraints (rationalization tables, drive-aligned framing) should STAY as prompt — hooks can't teach reasoning
Score based on: how many mechanical constraints are prompt-only when they could be hooks?

P21 — Auto-loader usage for constraints:

Do phase skills that load constraint prose use the bang-invoked auto-loader?

!`uv run python3 ${CLAUDE_SKILL_DIR}/../../scripts/load-constraints.py skill-name`

Or do they list Read() calls for each constraint .md file manually?
Why this matters: The auto-loader + applies-to frontmatter is the wiring that makes atomic constraints work. Manual Read() lists mean adding a new constraint requires editing every skill that should load it — silent drift is the default failure mode.
Check: Run uv run python3 ${CLAUDE_SKILL_DIR}/../../references/constraints/auto-loader-usage.py. Every flagged SKILL.md is a violation. (Bare relative path won't resolve from the auditor's CWD — always prefix with ${CLAUDE_SKILL_DIR}/../../.)
Exceptions: router skills that immediately delegate (no constraint evaluation), ad-hoc single-file references (not phase sets), plugins without scripts/load-constraints.py.
Score: count of phase skills using the loader / count of phase skills that load ≥2 constraints. Below 80% = critical gap.

Gate: Architecture Scored [checkpoint: human-verify, auto-advanceable]

Verify scores for all P01-P21 (+ P19b) principles are present
Each principle must have numeric score (0-10) + 1-line justification
If any principle ID is missing, score it now
Composite = average of scored (non-N/A) principles

Update .planning/wc/{name}/STATE.md:

step: 2-score
status: completed
implements: [WC-09]
requires: [all workflow skill files]
provides: [P01-P21 scores with justifications]
affects: [.planning/wc/{name}/STATE.md]
one-liner: "P01-P21 scored with line-number evidence by the wc-audit reviewers; composite computed in JS."

Proceed to Step 3. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 3: Score Against Enforcement Checklist

!cat ${CLAUDE_SKILL_DIR}/../../references/enforcement-checklist.md You MUST read this file before scoring. No scoring from memory.

For each of the 13 patterns, score:

Present - pattern exists and is well-implemented
Weak - pattern exists but is insufficient (e.g., soft language instead of Iron Law)
Absent - pattern is missing where it should exist

Identify the highest-drift phases with the weakest enforcement - these are the critical gaps.

Gate: Enforcement Scored [checkpoint: human-verify, auto-advanceable]

Verify all 13 patterns were scored
Each pattern must be marked: Present / Weak / Absent
If any pattern is missing, score it now

Update .planning/wc/{name}/STATE.md:

step: 3-enforcement
status: completed
requires: [enforcement-checklist.md, all workflow skill files]
provides: [13-pattern scores per phase]
affects: [.planning/wc/{name}/STATE.md]
one-liner: "13 enforcement patterns scored Present/Weak/Absent per phase; weakest high-drift phases flagged."

Proceed to Step 3b. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 3b: Audit Path Portability

Red Flags — STOP: declaring portability "Clean" without actually running the mandatory grep commands below · treating a non-zero ${CLAUDE_SKILL_DIR}-in-hook-command hit as a warning instead of a defect · assuming a path resolves because it "looks right" rather than tracing it from the user's CWD. (The April 2026 incident — 9 days of silently-broken hooks — was exactly a path that "looked right" and was never grep-audited.)

Skills run in the user's project CWD, not the plugin directory. Every path in a SKILL.md that references plugin-internal files must resolve regardless of CWD.

Scan every SKILL.md and references/*.md file in the workflow for these patterns:

Relative script paths — uv run python3 scripts/, uv run python3 ../, uv run python3 ../../ referencing plugin scripts
- These break because the agent's CWD is the user's project
- Fix: Use ${CLAUDE_SKILL_DIR}/../.. for absolute paths:
```
uv run python3 "${CLAUDE_SKILL_DIR}/../../skills/SKILL/scripts/script.py" args
```
- Or use ${CLAUDE_SKILL_DIR} for files within the same skill directory:
```
uv run python3 "${CLAUDE_SKILL_DIR}/scripts/script.py" args
```
Relative Read() paths — Read("../../skills/..."), Read("../audit-check/SKILL.md")
- The Read tool requires absolute paths; ../../ resolves from user's project CWD, not skill directory
- Fix: Use ${CLAUDE_SKILL_DIR}/../.. or ${CLAUDE_SKILL_DIR}:
```
Read `${CLAUDE_SKILL_DIR}/../../skills/SKILL-NAME/SKILL.md` and follow its instructions.
```
Dynamic context via bang-backtick injection — For constraint files that should be inlined at skill load time, use the pattern: exclamation mark followed by backtick-cat path backtick. Example: BANG + `cat ${CLAUDE_SKILL_DIR}/../../references/file.md`. This inlines the file contents at skill load time. Note: bang-backtick injection only works in top-level skills loaded via Skill(). Internal skills loaded via Read() should use direct Read() instructions instead.
Path variable substitution — the two variables apply to DIFFERENT contexts:

| Context | Correct Variable | Docs | |---------|------------------|------| | Hook command: fields in YAML frontmatter | ${CLAUDE_PLUGIN_ROOT} | hooks.md | | Skill content body (markdown + bash injection) | ${CLAUDE_SKILL_DIR} | skills.md | | Internal skills (loaded via Read) | Neither substitutes. Use ${CLAUDE_SKILL_DIR}/../../ as a convention so a consistent style is preserved — the agent infers the actual path from context. | — | | ${CLAUDE_SKILL_DIR} inside an Agent() prompt string | SAFE — it is skill-content body, so it is substituted to the literal absolute path at skill-load, before the orchestrator constructs the Agent() call. The spawned subagent receives the resolved path, never the token. No need to pre-resolve and pass it in. | — |

The hook/content variables ARE NOT INTERCHANGEABLE. ${CLAUDE_SKILL_DIR} is bound in the shell environment only when a skill is actively loaded via Skill(). When a hook fires for a tool call outside that active session — e.g., a matcher: "*" hook, or an Agent matcher spawned from main chat — the env var is empty and the path resolves to garbage like /../../hooks/foo.py.

Hook-command variable misuse — the April 2026 incident ⚠️
- What happened: course-materials plugin v2.83.1 (Apr 15 2026) switched hook commands from ${CLAUDE_PLUGIN_ROOT}/hooks/... to ${CLAUDE_SKILL_DIR}/../../hooks/... based on a misdiagnosis claiming CLAUDE_PLUGIN_ROOT was "unset at PreToolUse runtime". Five skills across two plugins (teaching + derivative exam skills) shipped with broken hook paths for 9 days.
- Why it stayed hidden: the affected hooks (no-agent-resume-guard.py, context-monitor.py) default to {"decision": "approve"} when the script runs. When the script doesn't exist at all, Claude Code also approves. The silent-failure mode was indistinguishable from a clean approval, so "no enforcement" looked like "approved every time".
- Why exam-prep finally exposed it: a new skill added matcher: "*" under PreToolUse for context-monitor.py. matcher: "*" fires on every tool call in the session, including tool calls before the new skill was ever invoked via Skill(). At that moment ${CLAUDE_SKILL_DIR} was empty, producing a nonexistent path — Claude Code blocked every tool call with the hook's failure.
- The lesson: ${CLAUDE_SKILL_DIR} in hook frontmatter is a silent-failure landmine. It appears to work because existing hooks default-approve. Add one blocking hook — or a broad matcher — and the whole plugin surfaces the latent bug at once.

Hook Command Variable Audit (mandatory during Path Portability review):

# This command should return EMPTY — any hit is a defect:
grep -rn "command:.*\${CLAUDE_SKILL_DIR}" skills/*/SKILL.md

# All hook commands should match this pattern:
grep -rn "command:.*uv run python3 \${CLAUDE_PLUGIN_ROOT}" skills/*/SKILL.md

If the first grep returns anything, flag as a Critical Gap.

Score:

Clean — no broken paths found AND no ${CLAUDE_SKILL_DIR} in hook command fields
Partial — some paths fixed, others remain
Broken — relative paths present in skill instructions OR ${CLAUDE_SKILL_DIR} in hook command fields (even if file paths happen to resolve when tested)

Gate: Path Portability Scored [checkpoint: human-verify, auto-advanceable]

Verify all SKILL.md and references/*.md files were scanned
Every uv run python3 ../ and Read("../ pattern was flagged
Score is recorded

Update .planning/wc/{name}/STATE.md:

step: 3b-portability
status: completed
requires: [all SKILL.md and references/*.md files]
provides: [path portability score]
affects: [.planning/wc/{name}/STATE.md]
one-liner: "Path portability scored Clean/Partial/Broken; hook-command variable audit run; candidacy scan fed into Step 4."

Ultracode-Workflow Candidacy Scan (feeds Step 4 Recommendations — no separate gate)

Before writing the report, scan every phase for ultracode-workflow migration candidates — fan-out phases that should be Claude Code ultracode workflows rather than in-skill agent dispatch. Read ${CLAUDE_SKILL_DIR}/references/dynamic-workflow-migration.md §1 for the rubric. Scan for BOTH worker modes — workflows are NOT read-only:

Review fan-out — N read-only agents (per section/lecture/question/source/footnote) → computed gate / structured findings.
Write/transform fan-out — N write-agents (per file/site/lecture/section) that create or transform artifacts from a fixed spec (codemod, migration, per-item spec-driven generation), worktree-isolated. The docs' flagship case: 500-file migration, "make the change." Do not skip these — they are often the strongest candidates.

Flag a phase when the SHAPE qualifies AND it wins on ≥1 value driver:

Shape (required): the phase dispatches N agents "one per X" over a known list, and either the skill consumes their aggregated results (review) OR each does an independent per-item mutation (write). A numeric gate is NOT required.
At least one value driver: (a) parallelism, (b) context isolation, (c) a deterministic gate replacing a model-reported "recompute by hand" score (strongest review signal), or (d) independent per-item mutation at scale (write fan-out — migrations, codemods, per-item generation).

The generation line — SPLIT it (do NOT blanket-leave generation): mechanical/spec-driven per-item creation/transformation (the "what" is pinned by an inventory/outline/rule — e.g. per-lecture slide creation from a 15-20-item inventory, per-section assembly from an outline, a codemod) → FLAG as a transform-workflow candidate. Only creative/judgment generation (brainstorm a thesis, draft novel prose where voice IS the work) stays conversational.

Not disqualifiers: a mid-run user strategy choice ("sequential or parallel?") stays in the skill. A diagnosis output (REVIEW.md) instead of a numeric score is fine (driver b). The phase writing files is NOT a disqualifier (that's exactly driver d).

For each flagged phase, classify strong / moderate, note worker-mode (review vs transform), and add a Recommendation: "Migrate <phase> to an ultracode workflow — review→gate/findings, or transform→worktree write + verify; keep the creative 'what' + /goal + R4 in the skill. Mode 3 improvement; see the migration playbook." Do NOT flag: single-agent phases (no fan-out), creative-judgment drafting/brainstorm, content-approval gates, routing, or fan-outs whose agents are external (Batch API / CLI, not Claude subagents). If a phase has a real fan-out but no value-driver win, note "fan-out present, no migration win". If nothing qualifies, state "no ultracode-workflow candidates" — silence is ambiguous.

Proceed to Step 4. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 4: Output Audit Report

Render AUDIT.md from the workflow result — do NOT re-score by hand. The wc-audit workflow already produced result.reportMarkdown (the full AUDIT.md body: P01-P21 table, enforcement coverage, path portability, candidacy table, critical gaps), result.scoreTable (the dimension-level gate), result.candidacyTable, result.composite, and result.verdict. Write result.reportMarkdown verbatim to .planning/wc/{name}/AUDIT.md and present result.scoreTable + result.composite to the user. The gate is result.overallPass/result.composite, computed in JS — do not recompute or rationalize it. Append the composite row to .planning/wc/{name}/SCORES.md for the Mode 3 trend.

The format below documents what result.reportMarkdown contains (so you can sanity-check the workflow's output) — it is the spec the workflow renders to, not a worksheet to fill in yourself.

Format:

## Audit: [Workflow Name]

### Architecture Scores (P01-P21)
| ID | Principle | Score | Notes |
|----|-----------|-------|-------|
| P01 | Phased decomposition | [0-10] | [notes] |
| P02 | Gates (deterministic/judgment) | [0-10] | [notes] |
| P03 | Structural gate enforcement | [0-10] | [notes] (STRUCTURAL/total) |
| P04 | Independent verification | [0-10] | [notes] |
| P05 | Artifact review (4-level) | [0-10] | [notes] |
| P06 | Two entry points | [0-10] | [notes] |
| P07 | Cross-skill consistency | [0-10] | [notes] |
| P08 | Constraint/convention coverage | [0-10] | [notes] |
| P09 | Iteration strategy | [0-10] | [notes] |
| P10 | Post-subagent enforcement | [0-10] | [notes] |
| P11 | Deviation rules | [0-10] | [notes] |
| P12 | State management | [0-10] | [notes] |
| P13 | Session handoff | [0-10] | [notes] |
| P14 | Checkpoint types | [0-10] | [notes] |
| P15 | Context monitoring | [0-10] | [notes] |
| P16 | Summary frontmatter | [0-10] | [notes] |
| P17 | Agent tool restrictions | [0-10] | [notes] |
| P18 | Requirement traceability | [0-10] | [notes] |
| P19 | Autonomous phase chaining | [0-10] | [notes] |
| P19b | Visual output | [0-10] | [notes] |
| P20 | Hooks over prompt | [0-10] | [notes] |
| P21 | Auto-loader usage | [0-10] | [notes] (loader skills / phase skills ≥2 constraints) |

### Gate Enforcement Matrix
| Transition | Gate | Artifact | Producer Writes? | Consumer Checks? | Hook Enforced? | Status |
|------------|------|----------|-------------------|-------------------|----------------|--------|
| [phase A] → [phase B] | [gate desc] | [artifact file] | ✅/❌ | ✅/❌ | ✅/❌ | HOOK/STRUCTURAL/ADVISORY ⚠️ |

### Enforcement Coverage
| Pattern | Phase 1 | Phase 2 | ... | Phase N |
|---------|---------|---------|-----|---------|
| Iron Laws | ✅/⚠️/❌ | ... | ... | ... |
| ... | ... | ... | ... | ... |

### Path Portability
| File | Pattern | Status |
|------|---------|--------|
| skills/X/SKILL.md | `uv run python3 scripts/foo.py` | ❌ Broken / ✅ Fixed |
| skills/Y/SKILL.md | `Read("../../lib/...")` | ❌ Broken / ✅ Fixed |

### Ultracode-Workflow Migration Candidates
| Phase | Fan-out? | Worker mode (review/transform) | Value driver | Recommend migrate? (strong/moderate) | Note |
|-------|----------|--------------------------------|--------------|--------------------------------------|------|
| [phase] | ✅/❌ (one per X) | review / transform / — | parallelism / context / gate / per-item-mutation | ✅ Mode 3 (strong\|moderate) / ❌ leave | [why] |

(From the Ultracode-Workflow Candidacy Scan. "no ultracode-workflow candidates" if none qualify.)

### Critical Gaps
1. [Highest priority gap + recommendation]
2. [Second priority gap + recommendation]
...

### Recommendations
[Specific, actionable changes]

Render score trend (if SCORES.md exists from a prior audit):

uv run python3 ${CLAUDE_SKILL_DIR}/../../scripts/render-audit-scores.py .planning/wc/{name}/SCORES.md

Traceability (self-applied P18): write or update .planning/wc/{name}/VALIDATION.md mapping each WC-NN requirement to the audit evidence that verifies it (the gate/principle that confirms it) + its scope tag — the same closure Mode 1 Step 7 performs, applied to an audit-only run.

Gate: Audit Reported [checkpoint: human-verify, auto-advanceable]

.planning/wc/{name}/AUDIT.md exists and contains result.reportMarkdown (P01-P21 table, enforcement coverage, path portability, candidacy table, critical gaps)
result.composite was appended to .planning/wc/{name}/SCORES.md
.planning/wc/{name}/VALIDATION.md maps the WC-NN evidence (P18 self-application)
The gate verdict presented to the user is result.overallPass verbatim — not recomputed

Persist audit results: Write result.reportMarkdown to .planning/wc/{name}/AUDIT.md in addition to displaying it. Append result.composite to .planning/wc/{name}/SCORES.md. Review-pattern logging (self-applied P19b): append to .planning/wc/{name}/LEARNINGS.md what the user attended to in the audit results (which findings they prioritized or overrode) — the same observe→record→offer loop Mode 1 Step 7 performs. Update .planning/wc/{name}/STATE.md:

step: 4-report
status: completed
implements: [WC-09]
requires: [all workflow skill files]
provides: [AUDIT.md, VALIDATION.md]
affects: [.planning/wc/{name}/AUDIT.md, .planning/wc/{name}/VALIDATION.md]
one-liner: "AUDIT.md rendered from the wc-audit result (composite + matrices + candidacy); VALIDATION.md maps WC-NN evidence."

<EXTREMELY-IMPORTANT> ### The Iron Law of Thorough Scoring

NO PRINCIPLE SCORE WITHOUT LINE-NUMBER EVIDENCE. This is not negotiable.

Mode 2 is the highest-drift mode: the auditor is tempted to skim, anchor to an overall impression, and give generous scores. Every principle score MUST cite specific line numbers or artifact names as evidence. A score without evidence is a guess, not an audit.

If you cannot point to a specific line, file, or pattern that justifies the score — the score is wrong. </EXTREMELY-IMPORTANT>

Deviation Rules for Mode 2 (Auditing)

During auditing, unplanned issues may arise. Apply these deviation rules:

| Rule | Trigger | Action | Permission | |------|---------|--------|------------| | R1: Broken path | Skill file path doesn't resolve, Read fails on referenced file | Note broken path in audit, score affected principles accordingly | Auto | | R2: Missing section | Expected section (gates, enforcement, state management) absent from workflow | Note as critical gap in audit report, don't invent content | Auto | | R3: Blocking format | Workflow file is malformed YAML, unparseable frontmatter, or encoding issue | Note the format issue, score what you can read, flag the rest | Auto | | R4: Scope change | Audit scope needs to expand (new files discovered, dependency chain found) | STOP — present the expanded scope to the user before continuing | Ask user |

Priority: R4 (STOP) > R1-R3 (auto) > unsure → R4

Mode 2 Enforcement

Rationalization Table — Mode 2

| Excuse | Reality | Do Instead | |--------|---------|------------| | "This workflow is simple, I can skim the skill files" | Skimming produces generous scores. You miss enforcement gaps in the middle of long files. March 2026: 285-line Step 3 hid 6 un-gated sub-responsibilities. | Read ALL files line by line. Use Read with offset/limit to chunk large files. | | "This principle obviously doesn't apply to this domain" | Score it anyway with a justification. The auditor's job is to score everything, not to pre-filter. A principle that "doesn't apply" gets a justified N/A, not a skip. | Score it. Write the justification. Let the reader decide if N/A is warranted. | | "The workflow has good enforcement overall, I'll give generous scores" | Generous audits ship broken workflows. April 2026: baseline 5.2 was generous at 6.5 before careful tally. | Score each principle independently. Sum at the end. Don't anchor to an overall impression. | | "Enforcement patterns don't matter for low-drift phases" | Low-drift phases still need gates. A brainstorm phase with no gate means the agent can skip directly to implementation. | Score enforcement for every phase. Note "low enforcement appropriate" in justification if warranted. | | "The audit is already long enough, I'll skip the matrices" | The Gate Enforcement Matrix and Hook Coverage Matrix catch gaps that prose misses. Without them, the audit is subjective. | Produce all required matrices. They take 5 minutes and catch what prose can't. |

Red Flags — STOP If You Catch Yourself:

| Action | Why Wrong | Do Instead | |--------|-----------|------------| | Scoring a principle without reading the relevant skill file section | You're scoring from memory or impression, not evidence | Read the specific section, find the line numbers, then score | | Giving a 9 or 10 without finding specific evidence | High scores require evidence of excellence, not absence of problems | Find the specific line/pattern that earns the score. No evidence = no high score | | Skipping the enforcement pattern matrix | "I covered it in prose" — matrices catch asymmetries prose misses | Produce the matrix. Score every pattern × every phase | | Scoring structural gate enforcement without producing the Gate Enforcement Matrix | You can't assess gates without mapping every transition | Produce the matrix first, then score from it | | Combining two principles into one score | Each principle measures a different quality dimension | Score each P01-P21 independently |

Staged Review — Mode 2

If the audit report composite is below 7.0 on first pass, STOP and re-read all workflow files from scratch before finalizing scores. Audits this low usually indicate the auditor missed a section or misunderstood the workflow structure — not that the workflow is truly that weak. Re-reading costs 5 minutes; a wrong baseline wastes the entire improvement cycle.

Delete & Restart — Mode 2

<EXTREMELY-IMPORTANT> **If you scored ANY principle before its evidence was available, DELETE all scores and restart from Step 1. No exceptions.** This fires in three cases: (1) you scored before reading the relevant file section; (2) you scored before the wc-audit workflow returned `result`; (3) you carried a score for a principle the reviewers did not actually cover this run. A partial-read or pre-result audit produces anchored scores that resist correction — starting fresh is faster than debiasing. </EXTREMELY-IMPORTANT>

Drive-Aligned Framing — Mode 2

| Your Drive | Why You Cut Corners | What Actually Happens | The Drive You Failed | |------------|--------------------|-----------------------|---------------------| | Helpfulness | "Quick audit so user gets results faster" | Generous audit ships broken workflow. User discovers gaps in production. | Anti-helpful | | Competence | "I can tell from the structure this is fine" | Without line-by-line reading, you miss the gap buried at line 285 of a 500-line file. | Incompetent | | Efficiency | "Matrices are overhead, prose covers it" | Prose catches narrative gaps. Matrices catch structural gaps. You shipped an asymmetric hook coverage. | Anti-efficient | | Approval | "The user wants the audit done quickly" | A fast, generous audit they later catch wrong costs you more trust than a slow honest one. They stop trusting your scores. | Lost approval | | Honesty | "I'll score this 8 because it's mostly there" | "Mostly there" means gaps exist. An honest score reflects the gaps. | Dishonest |

Mode 3: Improve Workflow

Migrating a fan-out phase to an ultracode workflow is a Mode 3 improvement. If the user asks to "migrate a phase to an ultracode workflow" / "convert fan-out to a workflow script," or the audit's candidacy scan flags a fan-out phase — review (read-only → gate/findings) OR transform (write-agents creating/transforming from a fixed spec, worktree-isolated; e.g. a codemod, a migration, per-item spec-driven generation) — treat the migration as the fix: read ${CLAUDE_SKILL_DIR}/references/dynamic-workflow-migration.md (decision rubric, both worker modes, the discover→transform→verify pattern for writes, script conventions, packaging, wiring, exit gate), confirm the candidate from the ACTUAL phase file (not a summary), write workflows/<name>.js, node --check it, verify the artifact lands at the expected path, then wire the skill (keep the creative "what" + /goal + R4). For migrating a backlog of several phases at once, prefer a one-off migration workflow (fan out over all candidates) rather than one-at-a-time.

<EXTREMELY-IMPORTANT> ## The Iron Law of Workflow Improvement

MODE 3 IS AN AUDIT-FIX LOOP. THE SUBSTRATE GATE DECIDES WHEN TO STOP, NOT YOU AND NOT A BARE COMPOSITE NUMBER. This is not negotiable.

Mode 3 terminates when result.substratePass is true AND the composite has gone FLAT (within ±0.2 of the prior iteration) — NOT when a noisy composite crosses an arbitrary 9.5. The substrate gate is the deterministic, monotonic signal: 0 critical findings · no enforcement pattern Absent where a phase needs it · path portability Clean. Those converge; the 0-10 composite does not.

The structural problem Mode 3 solves (two failure modes, opposite directions):

Stopping too early by self-declaring "diminishing returns" before the substrate is clean. The loop + /goal evaluator prevents this — you cannot stop while a critical or an Absent enforcement pattern remains.
Chasing a noisy proxy forever. The composite is an LLM panel: it re-rolls ±0.2 each run, its domain-ceiling denominator drifts, and it regenerates new minor findings every pass. Targeting composite ≥ 9.5 sends the loop onto a treadmill where fixing 4 findings surfaces 4 new ones and the number never converges (empirically every workflow asymptotes ~9.0; see project_wc_mode3_asymptote). The last 0.5 to 9.5 can only be bought by over-enforcing creative/low-drift phases — which violates the Step-4 drift tiering and the no-speculative-enforcement rule. So 9.5 rewards making the workflow worse. </EXTREMELY-IMPORTANT>

Why the gate is the substrate, not 9.5 (calibrated ceiling)

| | Composite (0-10) | Substrate gate | |---|---|---| | Stationary across runs? | No — ±0.2 LLM noise, drifting denominator | Yes — deterministic | | Converges as you fix? | No — regenerates findings (treadmill) | Yes — criticals 5→0, enforcement Weak→Present are monotonic | | Last-mile incentive | Push to 9.5 ⇒ over-enforce creative steps (anti-pattern) | Close real structural gaps only |

Doctrine: drive the loop to substratePass + composite ≥ 9.0 (calibrated ceiling) + composite flat across 2 iterations, then STOP. Do not iterate further to lift the composite — that is the treadmill. If substratePass is true but the composite sits at, say, 8.6 and is flat, the workflow is at its ceiling and sound: ship it and note the composite as the honest harsh-auditor reading, not a defect.

Step 1: Run Initial Audit (Mode 2)

State initialization: Create .planning/wc/{name}/STATE.md with this YAML template:

---
mode: improve
step: 1-initial-audit
status: in_progress
target: [workflow name]
implements: [WC-12]
requires: [target workflow files]
provides: [baseline composite]
affects: [.planning/wc/{name}/STATE.md, .planning/wc/{name}/SCORES.md]
one-liner: "Improve loop started on {target} — running baseline wc-audit before the /goal climb."
---

Context monitoring: Mode 3 runs multi-iteration audit-fix loops. Each iteration consumes significant context. Check availability:

If context is low (≤35% remaining), write .planning/wc/{name}/HANDOFF.md with current iteration, score, and remaining gaps before starting the next iteration.
If context is critical (≤25% remaining), write HANDOFF.md immediately — do not start another iteration.

Run Mode 2 on the target workflow. This produces the baseline score.

Gate: Mode 2 audit report exists with numeric scores for all P01-P21 principles. [checkpoint: human-verify, auto-advanceable]

Step 2: Launch Audit-Fix Loop

Use the audit-fix-loop pattern with /goal as the cross-turn iteration primitive (separate-model evaluator reads SCORES.md from the transcript). The canonical doctrine — auditor≠fixer, substrate-gate-not-bare-score, and the anti-grind rationale — lives in the audit-fix-loop skill; what follows is the wc-specific instantiation (wc-audit's substratePass / composite fields):

/goal Workflow [WORKFLOW_NAME] reaches result.substratePass=true (0 critical, no enforcement Absent, portability Clean) AND composite >= 9.0 AND composite flat (within ±0.2 of the prior SCORES.md row). Stop after 10 turns. Do NOT keep iterating to lift the composite once the substrate gate is clean and the composite is flat — that is the treadmill.

Before launching the goal, persist the second link of the improve chain (the hook's improve chain is 1-initial-audit → 1-audit-loop):

step: 1-audit-loop
status: in_progress
requires: [Mode 2 audit report]
provides: [score-gated fix iterations]
affects: [.planning/wc/{name}/STATE.md, .planning/wc/{name}/SCORES.md, target workflow files]
one-liner: "/goal audit-fix loop launched (gate: substratePass + composite≥9.0 + flat); each turn re-runs wc-audit and appends composite + substratePass to SCORES.md."

Each turn under the active goal follows this exact sequence:

Phase A: AUDIT ──→ Phase B: DECIDE ──→ Phase C: FIX
  [fresh subagent]   [check score]       [targeted edits]
    │                    │                    │
    ▼                    ▼                    ▼
  AUDIT.md         substratePass &&      Fix gaps by priority:
  SCORES.md        composite≥9.0 &&       1. criticals (any severity:critical)
                   flat(±0.2)?            2. enforcement Absent → Present
                        │                 3. portability → Clean
                    YES ──→ end turn;      4. then P<9 medium gaps
                           /goal evaluator      │
                           marks done           ▼
                        │                   End turn → /goal
                    NO ──→ Phase C          refires Phase A

Phase A: AUDIT (the wc-audit workflow — MANDATORY, independent by construction)

Run the Mode 2 wc-audit ultracode workflow (see "How Mode 2 runs" above) — it IS the fresh, independent audit. Each dimension reviewer reads the files cold with NO knowledge of your fixes, and the composite is computed in JS, so a fixer cannot rubber-stamp its own work:

WF=$(command ls -d ~/.claude/plugins/cache/edwinhu-plugins/workflows/*/workflows/wc-audit.js 2>/dev/null | sort -V | tail -1)
[ -z "$WF" ] && WF="${CLAUDE_SKILL_DIR}/../../workflows/wc-audit.js"

# Full audit on iteration 1; selective re-audit thereafter:
Workflow({ scriptPath: "<WF>", args: {
  targetWorkflow: "{name}", projectDir: "<abs repo root>", pluginRoot: "<abs .../workflows dir>",
  threshold: 9.0,                                   // calibrated ceiling, NOT 9.5 — composite is advisory; substratePass is the gate
  onlyChecks: <prev result.reviewersThatFlagged>,   // omit on iteration 1
  priorReviews: <prev result.reviews>               // omit on iteration 1
} })

Write result.reportMarkdown to .planning/wc/{name}/AUDIT.md and append a row to .planning/wc/{name}/SCORES.md recording result.composite, result.substratePass, result.substrate (criticals / enforcementAbsent / portability). The workflow's reviewers are read-only (they REPORT, never fix) and the gate is computed in JS. result.substratePass is the trustworthy convergence signal; result.composite is an advisory ±0.2 LLM proxy — record it, but the gate keys on substratePass + flatness, not the bare number.

<EXTREMELY-IMPORTANT> **THE AUDIT IS THE wc-audit WORKFLOW, NOT YOUR OWN RE-READ. If you score your own fixes by hand, you are rubber-stamping.**

The workflow's reviewers have no context from the fix phase — they read the files cold and the composite is computed in JS from raw scores. This is what makes the score trustworthy. Do not substitute a hand audit; do not recompute result.composite. </EXTREMELY-IMPORTANT>

Phase B: DECIDE `[checkpoint: decision]`

Read .planning/wc/{name}/SCORES.md. Render the score trend for visual context:

uv run python3 ${CLAUDE_SKILL_DIR}/../../scripts/render-audit-scores.py .planning/wc/{name}/SCORES.md

Check the substrate gate, then the calibrated composite + flatness. "Flat" = this iteration's composite is within ±0.2 of the prior SCORES.md row.

| Condition | Action | |-----------|--------| | result.substratePass AND composite ≥ 9.0 AND flat | Output <promise>WORKFLOW_DONE</promise> — substrate clean at the calibrated ceiling; STOP (do not iterate to lift the composite) | | result.substratePass AND composite ≥ 9.0 but NOT yet flat | Continue to Phase C only for any remaining sub-9 medium gaps; if none are cheap/real, treat as flat and STOP | | NOT result.substratePass (a critical, an enforcement Absent, or portability not Clean) AND iteration < 10 | Continue to Phase C — fix the substrate blocker (this is the non-negotiable part) | | NOT result.substratePass AND iteration ≥ 10 | Escalate to user with the substrate blockers and scoreTable | | result.substratePass but composite < 9.0 AND flat across 2 iterations | Escalate to user: substrate is clean but the composite sits below the ceiling and won't climb — likely a genuine domain ceiling or a judge artifact; present it, don't grind |

You may output the completion promise once result.substratePass is true and the composite is at/above the calibrated ceiling and flat. You may NOT keep iterating to push a flat composite toward 9.5 — that is the treadmill the Iron Law forbids. You may NOT declare done while substratePass is false (a critical / Absent enforcement / dirty portability remains) — the substrate is non-negotiable.

Review-pattern logging (self-applied P19b): append to .planning/wc/{name}/LEARNINGS.md after each DECIDE: which principle moved this iteration, the cheapest fix that moved it, and (if the user interjected) what they attended to. After 3+ iterations with a recurring drag-principle, propose encoding its fix as a default in the skill.

Phase C: FIX

<EXTREMELY-IMPORTANT> **NO FIX WITHOUT A SPECIFIC AUDIT FINDING. Targeted changes only — never rewrite a phase file wholesale.** Each edit must close ONE gap the wc-audit workflow named. Editing files the audit did not flag is unverified change that the next audit cannot attribute — and risks regressing a principle that was passing.

Delete & Restart (Phase C): if you have already edited a target file WITHOUT a corresponding AUDIT.md finding — revert that edit, find the finding that actually justifies the change (or drop it), then re-apply as a targeted fix. Fixes applied outside the audit are unverified and routinely regress a passing principle. (This run's own lesson: labeling the review markers "structural" without the enforcing hook regressed P03 and P20 — the hook had to follow the prose.) </EXTREMELY-IMPORTANT>

Post-subagent boundary (after wc-audit returns). The audit is the wc-audit workflow's job; fixing is yours. Keep the two separate:

| Main chat CAN do (fixing) | Main chat CANNOT do (re-auditing / investigation) | |----------------------------|-----------------------------------------------------| | Read AUDIT.md / SCORES.md and the specific files a finding cites | Re-score a principle by hand (the JS gate + reviewers own the composite) | | Edit the target files to close a named finding | Override result.composite / result.overallPass ("the auditor was wrong") | | Re-run the wc-audit workflow (full or onlyChecks) | Declare "close enough" below threshold | | git/ls/node --check to confirm an edit landed | Bash/Grep-spelunk the workflow to form your own score instead of re-running the audit |

After fixing, you re-run the workflow — you do NOT substitute your own read for the auditor's verdict.

Topic-change protocol (the /goal loop is iterative — an off-topic message must not silently kill it). If the user interjects mid-loop: (1) announce "Pausing the wc audit-fix loop at iteration N (composite X) to handle your request"; (2) handle the request; (3) announce-resume "Resuming the loop from iteration N" and re-fire Phase A. Never abandon the loop without saying so.

Address findings from .planning/wc/{name}/AUDIT.md, prioritized by severity:

Fix all principles scoring < 7.0 first (critical gaps)
Then principles scoring 7.0-8.9 (medium gaps)
Then principles scoring 9.0-9.4 (polish)

Fix reference — common gap → fix mapping:

| Gap | Fix | |-----|-----| | Missing Iron Law | Write with <EXTREMELY-IMPORTANT> tags | | Missing Rationalization Table | 5-10 entries (Excuse → Reality → Do Instead) | | Weak gate | Replace with verifiable condition | | Self-review as final gate | Add fresh subagent reviewer dispatch | | Missing Red Flags | 3-5 wrong-path indicators | | Missing Drive-Aligned Framing | 5-drive table (helpfulness > competence > efficiency > approval > honesty) | | No shared enforcement across skill family | Move rules to references/constraints/; all domain skills Read() the specific .md files they need | | Rules in separate constraints/ and conventions/ directories | Merge into single constraints/ directory. Presence of .py file = constraint; .md only = convention | | Monolithic shared constraints file >15 sections | Refactor to atomic files in constraints/ — one .md per rule, co-located .py for testable rules | | Constraint .md without co-located .py check script | Write the .py file alongside the .md. Same name, same directory. Auto-discovered by runner | | No auto-discovering test runner | Create check-all.py that globs constraints/*.py — no manual wiring needed | | Manual test runner that lists scripts explicitly | Replace with auto-discovery (glob("constraints/*.py")). Adding a .py file = automatically tested | | Verification only runs scripts OR only does LLM review | Both legs: check-all.py (hard block on .py failures) + reviewer subagent (soft block, scores .md-only conventions) | | Convention that could be a constraint | Graduate it: write the .py check script alongside the existing .md. No file moves needed | | check-script frontmatter linking .md to .py | Remove — naming convention handles the link. foo.md + foo.py = paired | | Index/TOC files manually listing constraints | Remove — the filesystem IS the index. ls constraints/*.md = all rules, ls constraints/*.py = all tests | | Hooks inconsistent across skill family | Produce Hook Coverage Matrix (skills × hooks); add missing hooks to skill frontmatter; justify intentional gaps | | Constraint added to individual skill but applies to family | Move to references/constraints/ with applies-to frontmatter; remove from individual skill | | Missing artifact review gate | Add reviewer subagent dispatch between producing/consuming phases, /goal-driven (5-turn budget; evaluator gates exit on reviewer APPROVED) | | Nested agent dispatch (agent spawns sub-agents) | Flatten: orchestrator spawns all agents directly in parallel. Move "dispatcher" logic into the skill definition. | | Broken paths (script) | Use ${CLAUDE_SKILL_DIR}/../../skills/SKILL/scripts/script.py | | Broken paths (Read) | Use ${CLAUDE_SKILL_DIR}/../../skills/SKILL-NAME/SKILL.md | | Missing post-subagent enforcement | Add verification/investigation boundary table for the domain | | Hook command uses ${CLAUDE_SKILL_DIR} instead of ${CLAUDE_PLUGIN_ROOT} | Replace with ${CLAUDE_PLUGIN_ROOT}/hooks/script.py. ${CLAUDE_SKILL_DIR} only substitutes in skill content, not hook frontmatter — it's empty when hooks fire outside an active Skill() session, silently failing. See April 2026 incident (teaching plugin v2.83.1→v2.84.4). | | Hook matcher: "*" under PreToolUse without hardened script | Move to PostToolUse unless the hook genuinely needs to block every tool call. A PreToolUse "*" hook that fails (bad path, missing file, non-zero exit) blocks every tool in the session — so the hook script must be resilient (defaults to approve on error) and its command path must be rock-solid. | | Missing topic change protocol | Add announce-pause / handle / announce-resume | | Missing deviation rules | Add 4-rule system (R1-R3 auto, R4 STOP) adapted to domain | | Missing state folder | Consolidate into .planning/ with standard files | | Missing session handoff | Add .planning/HANDOFF.md check to entry point startup | | Missing checkpoint types | Classify every gate as human-verify/decision/human-action | | Missing context monitoring | Add thresholds: warning ≤35%, critical ≤25% | | Missing summary frontmatter | Add YAML frontmatter with implements/requires/provides/affects | | Missing agent tool restrictions | Add allowed-tools to reviewer/verifier skills | | Missing requirement traceability | Add CATEGORY-NN IDs in spec, trace through plan and validation | | Missing autonomous phase chaining | Add auto-advance for human-verify gates, smart-discuss batching | | Mechanical constraints enforced only via prompt | Write scoped PreToolUse/PostToolUse hooks in skill frontmatter. File extension guards, path guards, tool param validation, sequence enforcement → hooks. Keep rationalization tables, drive-aligned framing, and quality judgments as prompt text |

Deviation Rules for Mode 3 Phase C (Fixing)

During fix application, unplanned issues may arise. Apply these deviation rules:

| Rule | Trigger | Action | Permission | |------|---------|--------|------------| | R1: Fix regression | A fix for one principle breaks another (e.g., adding a hook causes a constraint script false positive) | Revert the specific fix, note the regression in STATE.md, try alternative approach | Auto | | R2: Fix incomplete | Fix partially addresses the gap but can't fully close it in one edit | Apply partial fix, note remaining work in AUDIT.md for next iteration | Auto | | R3: Blocking dependency | Fix requires a file/hook/script that doesn't exist yet | Create the dependency first, then apply the fix, track both in STATE.md | Auto | | R4: Approach change | Fix requires restructuring the skill file, splitting phases, or changing the architecture | STOP — present the approach change to the user before proceeding | Ask user |

Priority: R4 (STOP) > R1-R3 (auto) > unsure → R4

Fix rules:

Targeted changes only — do NOT rewrite entire skill files
Each fix addresses ONE gap from the audit
After fixing, do NOT self-assess — the next iteration's audit will judge
End your turn immediately so the loop feeds you back for re-audit. Do NOT ask "should I continue?", do NOT summarize the fixes you just made, do NOT wait for confirmation — the /goal evaluator re-fires Phase A on its own. Pausing between fix iterations is procrastination disguised as courtesy; it strands the loop.

Efficiency Optimizations

Mode 3 can be expensive. These optimizations reduce cost without sacrificing audit independence.

1. Batch Fixes Per File

Group all fixes targeting the same file into a single edit. Don't make 5 separate edits to the same constraint file — read the audit findings, plan all changes, apply them in one Edit call.

2. Scoped Re-Audit After Iteration 1

The first audit (baseline) must read ALL files — no shortcuts. After that, subsequent audits can be scoped: the audit subagent reads all files but the fix agent only needs to re-read files it changed + the specific constraint .md files relevant to the fix. The audit subagent always does a full read (independence requires it), but the fixer can be smarter about what it reads before fixing.

3. Prioritize Cheapest High-Impact Fixes

After the audit, sort gaps by impact / effort:

Adding a .md + .py pair to constraints/ (all skills inherit, auto-discovered) = high impact, low effort
Adding allowed-tools frontmatter to 3 reviewer skills = high impact, 5 seconds each
Rewriting a phase skill's entire gate structure = medium impact, high effort

Fix the cheap high-impact gaps first. This maximizes score improvement per iteration.

4. Domain-Appropriate Scoring

Some principles have natural ceilings in certain domains:

Writing gates are judgment-based (not deterministic) — 9.0 is the natural ceiling for "gates" in writing
Writing has one midpoint because the domain only needs one — 9.0 is appropriate for "two entry points"

The auditor should note when a score reflects a domain ceiling vs a fixable gap. Domain ceilings don't count against the composite if the auditor justifies them. The composite then averages only the non-ceiling scores.

Caution: This is the auditor's call, not the fixer's. The fixer cannot declare a domain ceiling to avoid work. Only the independent auditor can classify a principle as domain-limited.

Why This Must Be a Loop (Not Manual Iteration)

The old Mode 3 had a flowchart showing a loop but no loop infrastructure. It relied on the agent manually deciding to continue iterating. This is an honor system — the exact failure mode that audit-fix-loop was built to prevent.

What happened (March 19, 2026): Agent stopped at 8.5 with a critical still open, self-rationalizing "diminishing returns." The failure wasn't "stopped below a magic number" — it was stopping while the substrate gate was still failing (a critical outstanding) and without a /goal loop forcing a re-audit. The /goal loop + substratePass gate prevent exactly that: you cannot stop while a critical / Absent enforcement / dirty portability remains.

The structural fix: /goal drives the iteration; a separate evaluator reads .planning/SCORES.md and marks done only when substratePass is true AND the composite is flat at/above the calibrated ceiling. This cuts BOTH failure modes: it won't let you stop with a substrate blocker, and it won't make you grind a flat composite toward an unreachable 9.5.

Rationalization Table — Mode 3

| Excuse | Reality | Do Instead | |--------|---------|------------| | "8.6 is below 9.5, keep grinding" | If substratePass is true and the composite is FLAT, 9.5 is unreachable noise — grinding it forces over-enforcement of creative steps (the anti-pattern). | Stop. Ship at the calibrated ceiling. Record the composite as the honest harsh reading. | | "substratePass is true so I'm done" (composite still climbing, not flat) | Not yet — a non-flat composite means cheap real gaps may remain. | One more pass on sub-9 medium gaps; stop once flat or no cheap real fix remains. | | "The critical is a domain characteristic / measurement artifact" | A critical keeps substratePass false — you cannot stop on it. But verify it first: a worktree's ambient .planning/ SPEC can make traceability checks fire spuriously (see asymptote memory). | Confirm the critical is real against the actual files; fix it or, if it's a confirmed measurement artifact, neutralize the artifact — don't just ignore it. | | "I'll run the audit manually instead of using /goal" | Manual loops have no enforcement; you'll stop early. | Set /goal pinned to substratePass + flat. The evaluator decides. | | "Re-reading all files each iteration is wasteful" | Fresh-context audit is what makes the substrate trustworthy. | Full re-read every iteration (selective via onlyChecks after iter 1). |

Red Flags — STOP If You Catch Yourself:

| Action | Why Wrong | Do Instead | |--------|-----------|------------| | Iterating to push a FLAT composite from ~9.0 toward 9.5 | The treadmill — the judge regenerates findings and the last mile rewards over-enforcement. | Stop once substratePass + flat. The composite ceiling is ~9.0, not 9.5. | | Declaring done with substratePass false | A critical / Absent enforcement / dirty portability is a real structural gap, not noise. | Keep iterating or escalate at the budget. | | Running Mode 3 without /goal | Honor system — you'll stop early. | Set /goal pinned to substratePass + composite≥9.0 + flat. | | Auditing your own fixes by hand | Rubber-stamping. | Re-run the wc-audit workflow (read-only reviewers, JS gate). | | Treating the composite as the gate | It's a ±0.2 LLM proxy with a drifting denominator. | The gate is substratePass; the composite is advisory + the flatness check. |

Drive-Aligned Framing — Mode 3

| Your Drive | Why You Cut Corners | What Actually Happens | The Drive You Failed | |------------|--------------------|-----------------------|---------------------| | Helpfulness | "Ship the fixes faster, skip the re-audit" | Unverified fixes ship broken enforcement; the user discovers a substrate gap in production. | Anti-helpful | | Competence | "I can tell from the diff this fixes the gap" | Self-assessment is rubber-stamping; the auditor exists because you can't see your blind spots. | Incompetent | | Efficiency | "Re-reading all files each iteration is wasteful" | Fresh-context audit catches regressions incremental review misses. | Anti-efficient | | Approval | "I'll grind to 9.5 because a bigger number looks better to the user" | You burn tokens on a treadmill and over-enforce creative steps, making the workflow worse. A flat 8.7 with substratePass beats a gamed 9.5. | Lost approval | | Honesty | "I'll call it 9.5-equivalent" or "I'll ignore the open critical" | Misrepresenting a noisy proxy, or shipping with a real substrate blocker, is fabricating completion. | Dishonest |

<EXTREMELY-IMPORTANT> ## Iron Laws of Workflow Creation

NO WORKFLOW WITHOUT PHILOSOPHY

Every workflow must trace back to PHILOSOPHY.md. If you can't explain how a phase serves phased decomposition, gates, or adversarial review, the phase doesn't belong.

NO PHASE WITHOUT A GATE

Every phase needs a gate — deterministic (test passes, file exists) or judgment-based (agent/human evaluates quality). Use the strongest gate available for the domain. No gate = not a real phase.

NO HIGH-DRIFT PHASE WITHOUT ENFORCEMENT

Identify where the agent is most tempted to shortcut. Enforce hardest there. Implementation and verification phases ALWAYS need Iron Laws.

NO UNREVIEWED ARTIFACT CROSSING A PHASE BOUNDARY

If a phase produces an artifact (spec, plan, outline) that downstream phases consume, the artifact MUST be independently reviewed before the next phase starts. Self-review is rubber-stamping. A fresh subagent reviewer catches what the author cannot see.

NO SKILL FAMILY WITHOUT SHARED ENFORCEMENT

If multiple skills in the same plugin operate on the same domain, their common enforcement MUST be consistent across THREE layers: (1) shared constraints file that every skill Read()s, (2) identical hooks in every skill's YAML frontmatter (or justified gaps), (3) every check script wired into the batch orchestrator, hook frontmatter, AND check definitions. Without three-layer consistency, skills enforce different rules — and the user gets different quality depending on which skill they invoke.

The course-materials incident (March 2026): batch-check-guard was added to slides-edit but not lecture-prep. convention-check-guard was added to sub-agent prompts but not as a hook. check-conventions.py existed but wasn't in check-all.sh. Result: lecture-prep shipped work that slides-edit would have caught. The constraints file was shared, but hooks and script wiring were not — two of three layers failed silently.

NO CONSTRAINT WITHOUT A CO-LOCATED CHECK SCRIPT

A constraint is a mechanically testable rule. Every constraint .md MUST have a co-located .py check script with the same name in the same constraints/ directory. No frontmatter linking, no manual wiring — same name = paired. The auto-discovering runner (check-all.py) globs constraints/*.py. If your constraint doesn't have a .py file, it's a convention (judgment-only), not a constraint. A .md that claims to be a constraint but has no .py is an untested unit test.

NO VERIFICATION WITHOUT BOTH LEGS

Verification has two legs: (1) constraint checks via auto-discovering check-all.py — runs all constraints/*.py files, structured JSON output, hard block on failure; (2) convention scoring via reviewer subagent — loads the .md-only files (no .py pair = convention), scores work against them, soft block below threshold. Running only one leg is incomplete. Scripts catch mechanical violations LLMs miss. LLMs catch quality issues scripts can't test. Both are necessary, neither alone is sufficient.

NO VERIFIER WITH WRITE ACCESS

Verification and review agents MUST use allowed-tools frontmatter restricting them to read-only tools. A verifier that can Write/Edit will "fix" issues it finds — silently bypassing the plan-execute-verify cycle. The fix was never planned, never reviewed, never tested. Tool restrictions make verification structurally honest, not just procedurally independent.

NO LONG WORKFLOW WITHOUT CONTEXT MONITORING

Workflows with 4+ phases MUST plan for context exhaustion. Warning at ≤35% remaining context (complete current task, then handoff). Critical at ≤25% (immediate handoff). An agent that starts a 10-task implementation phase with 20% context remaining will produce garbage for the last 5 tasks.

NO NESTED AGENT DISPATCH (Iron Law of Flat Dispatch)

Never design a workflow where an agent spawns its own sub-agents. The orchestrator (main chat or phase skill) MUST spawn all agents directly in parallel. Three-layer delegation (orchestrator → dispatcher agent → sub-sub-agents) fails because sub-sub-agent results don't reliably return via SendMessage — the middle dispatcher times out or loses results.

The course-materials incident (March 2026): slides-edit spawned teaching:reviewer, which dispatched 5 background sub-sub-agents (slide-auditor, notes-auditor). The reviewer returned without a final report. Had to be called 3 times — the 3rd time with "do NOT spawn sub-agents, run ALL checks inline." Fix: slides-edit now spawns 4-5 review agents directly. All return reliably.

BAD:  orchestrator → dispatcher agent → 5× sub-sub-agents (results lost)
GOOD: orchestrator → 5× agents directly in parallel (all return reliably)

When an agent needs multiple checks: The orchestrator reads the check list and spawns each check as a direct parallel agent. The "dispatcher" logic lives in the skill/phase definition, not in a middle agent.

The structural fix — ultracode workflows: For a genuine fan-out (one reviewer per item × check) producing a computed gate, migrate the dispatch into a Claude Code ultracode workflow (see ${CLAUDE_SKILL_DIR}/references/dynamic-workflow-migration.md; build it during Mode 1 decomposition or migrate an existing phase via Mode 3). An ultracode workflow is a script, not a dispatcher agent: reviewer results land in script variables and the gate is computed in JS, so result loss is impossible by construction — and the model can no longer inflate a self-reported score. Use it when a phase fans out + gates; keep drafting, /goal, and user-input phases conversational in the skill. </EXTREMELY-IMPORTANT>

Red Flags - STOP If You Catch Yourself:

| Action | Why Wrong | Do Instead | |---|---|---| | Creating a workflow without reading PHILOSOPHY.md | You'll miss the foundational principles | Read it first, every time | | Skipping the user interview | You'll design for an imagined domain, not the real one | Ask the six questions | | Writing soft language instead of Iron Laws | LLMs ignore polite suggestions | Use strong framing with EXTREMELY-IMPORTANT tags | | Proposing ungated phase transitions | Quality will die at the ungated boundary | Define a verifiable gate condition | | Designing all phases with equal enforcement | Drift risk varies by phase | Score enforcement density per phase | | Creating domain skills without shared enforcement | Each skill enforces its own version of the rules. lecture-prep misses checks that slides-edit catches — user has to run multiple skills to get consistent quality. | Co-locate common rules in references/constraints/ — skills Read() the specific .md files they need | | Adding a hook to one skill without checking siblings | Hook fires in slides-edit but not lecture-prep. The user gets different enforcement depending on which skill they invoke. Silent — no error, just missing enforcement. | Produce Hook Coverage Matrix. Add hook to all relevant siblings or justify the gap. | | Adding a constraint to an individual skill that should be shared | Constraint works in lecture-prep but notes-edit doesn't have it. User discovers the gap when notes-edit ships work that lecture-prep would have caught. | Add .md + .py to references/constraints/ with applies-to frontmatter. Over-inclusion beats drift. | | Growing a monolithic constraints file past 20+ sections | Every skill loads 450+ lines of constraints when it needs 5. Context bloat, slow loads, hard to maintain. | Refactor to atomic files in constraints/. One .md per rule, co-located .py for testable ones. Skills Read() only what they need. | | Writing a constraint .md without a co-located .py | A constraint without a test is like a unit test without an assertion — it documents intent but verifies nothing. | Write the .py alongside the .md. Same name, same directory. Auto-discovered by runner. | | Manually wiring check scripts into a runner | Manual wiring means new checks silently fail to run. Adding a file should be enough. | Use auto-discovering runner (glob("constraints/*.py")). No registration needed. | | Verification phase that only runs scripts OR only does LLM review | Scripts catch mechanical violations but miss quality. LLM review catches quality but misses mechanical violations. | Run both legs: check-all.py (hard block) + reviewer subagent scoring .md-only conventions (soft block). | | Letting an artifact pass to the next phase without review | Bad specs become bad designs become bad implementations. A 30-second review saves hours. | Add artifact review gate between producing and consuming phases | | No enforcement at the post-subagent boundary | That's where 71 violations happened in dev-debug (March 16). Main chat "verifies" by investigating. | Define verification/investigation boundary explicitly for the domain | | No topic change protocol in iterative loops | Off-topic user messages silently kill the loop. User has to re-invoke the skill. | Add announce-pause / handle / announce-resume protocol | | Rationalizations are hypothetical, not grounded | "Agents sometimes skip" is ignorable. "March 16: 71 violations, 3 re-invocations" is not. | Cite real failed sessions with dates, IDs, and violation counts | | Implementation phase with no deviation rules | Agents encounter unplanned work and either silently change architecture or halt on trivial bugs. | Add 4-rule deviation system with auto-fix for R1-R3, STOP for R4 | | State files scattered across .claude/ and project root | Next session can't find state; handoff fails. | Consolidate into .planning/ directory | | No handoff support in entry points | Context window exhaustion means lost work — next session starts from scratch. | Check for .planning/HANDOFF.md at startup, support structured resume | | Verification agent with Write/Edit access | Verifier silently "fixes" issues, bypassing plan-execute-verify. The fix was never planned or tested. | Add allowed-tools frontmatter restricting to Read, Grep, Glob only | | All gates treated as human-required | Workflow stops 7 times for rubber-stamp approvals. Unusable in autonomous/overnight mode. | Classify gates: human-verify (auto-advance), decision (pause), human-action (manual) | | No context monitoring in multi-phase workflow | Agent starts expensive phase with 20% context, produces degraded output, loses state. | Add context checks at phase entry, trigger handoff at ≤35% | | Phase summaries are unstructured prose | Handoff/resume requires re-reading all files. No dependency graph for parallel execution. | Add YAML frontmatter with implements/requires/provides/affects | | Requirements have no unique IDs | "We tested auth" doesn't tell you if login, refresh, AND logout are covered. | Assign IDs in .planning/SPEC.md, trace through .planning/PLAN.md and .planning/VALIDATION.md | | Every phase requires manual invocation | 7-phase workflow needs 7 human interventions to run. | Add autonomous chaining with auto-advance for human-verify gates | | Decision checkpoint with no review pattern tracking | You don't know what the human looks at, so you can't optimize for it. | Log what the human asks for at each review. After 3+ patterns, offer to automate. | | Designing an agent that spawns its own sub-agents | 3-layer delegation fails — sub-sub-agent results don't reliably return via SendMessage. The middle dispatcher times out or loses results. March 2026: teaching:reviewer dispatched 5 sub-agents, returned empty 2/3 times. | Use flat parallel dispatch: the orchestrator spawns ALL agents directly. Put the "dispatcher" logic in the skill definition, not in an agent. | | Copying a hook frontmatter pattern from a sibling skill without checking the variable | Sibling skills can carry latent bugs that only surface under specific matcher combinations. The teaching plugin carried a ${CLAUDE_SKILL_DIR} vs ${CLAUDE_PLUGIN_ROOT} bug in hook commands for 9 days (v2.83.1–v2.84.4) — silently default-approving because hook scripts also default to approve. Adding a single matcher: "*" hook in a new skill exposed the latent bug across the plugin. | Check docs: ${CLAUDE_PLUGIN_ROOT} for hook command: fields, ${CLAUDE_SKILL_DIR} for skill content. Never trust a sibling's pattern without verifying. | | Using matcher: "*" under PreToolUse without a bulletproof hook script | matcher: "*" fires for every tool call in the entire session — including calls before the skill was loaded. A hook script that can't find its path or exits non-zero blocks every tool. | Move to PostToolUse unless blocking is genuinely needed. If blocking IS needed, use a specific matcher (Write\|Edit\|Agent) and a defensively-coded script. |

Rationalization Table

| Excuse | Reality | Do Instead | |---|---|---| | "This workflow is simple, doesn't need enforcement" | Simple workflows drift fastest because the agent thinks it can shortcut | Add enforcement proportional to drift risk | | "Iron Laws feel too aggressive" | LLMs ignore polite suggestions. Strong framing works. | Write the Iron Law. It will be ignored if weakened. | | "Not every phase needs a gate" | Ungated phases are where quality dies | Define a verifiable gate condition | | "The user will catch errors in review" | Relying on human review defeats the purpose of the workflow | Build adversarial review INTO the workflow | | "I'll add enforcement later" | Later never comes. Enforcement debt compounds. | Add it now, refine through use | | "This domain is different, dev patterns don't apply" | The three pillars are universal. Enforcement density varies, principles don't. | Apply pillars, adjust density | | "Each skill can have its own enforcement" | Then lecture-prep misses what slides-edit catches, and the user runs 3 skills to get what 1 should provide. | Shared enforcement file. One source of truth for the domain. | | "This hook only applies to slides-edit, not lecture-prep" | Hooks enforce mechanical constraints. If the constraint applies to the domain, it applies to ALL skills in the domain. A hook that fires in one skill but not its sibling creates inconsistent enforcement with no visible warning. | Add to all sibling skills or justify in the Hook Coverage Matrix. | | "I'll add this constraint to the shared file later" | Later never comes. The rule lives in one skill, others ship without it. | Add .md (+ .py if testable) to constraints/ NOW. Over-inclusion is cheaper than drift. | | "One big constraints file is simpler than many small files" | Simpler to create, harder to maintain. At 20+ sections, every skill loads 400+ lines it doesn't need. | Atomic files: one .md per rule, co-located .py for testable ones. | | "I need an index file to track all constraints" | The filesystem IS the index. ls constraints/*.md = all rules. ls constraints/*.py = all tests. | No index needed. Auto-discovery replaces manual tracking. | | "This rule doesn't need a check script" | If it's mechanically testable, it needs a .py file. If it can't be tested, it's a convention — that's fine, just don't call it a constraint. | Ask: "Can I write a script that returns pass/fail?" If yes, write the .py. | | "I'll write the check script later" | Later never comes. The .md ships without its .py. The runner auto-discovers nothing. | Write .md and .py together. They're a pair. | | "I need to register the check script in the runner" | Auto-discovery means adding the .py file IS registration. Manual wiring is a bug source. | Just add the .py file. The runner globs constraints/*.py. | | "The LLM review covers everything, scripts are redundant" | LLM review drifts, rationalizes, and misses mechanical violations. Scripts are deterministic. | Scripts for mechanical checks, LLM for judgment. Both legs of verification. | | "This convention could be a constraint but testing it is hard" | Start it as a convention. Note it as a graduation candidate. Revisit when you have a testing idea. Don't force-fit a bad test. | Classify honestly. Graduation happens when testability improves, not when you want fewer conventions. | | "The spec looks fine, no need to review it" | Self-review is rubber-stamping. The author can't see their own blind spots. | Dispatch a fresh reviewer subagent. 30 seconds saves hours. | | "Plan review will slow us down" | A bad plan costs 10x more to fix during implementation than during review. | Review the plan. Fix it now, not during implementation. | | "The reviewer can just fix small issues it finds" | That bypasses plan-execute-verify. The "fix" was never planned, never reviewed, never tested. Now you have unverified code in production. | Restrict verifiers to read-only tools. Issues go back to the executor. | | "Context monitoring is overkill for short workflows" | A 4-phase workflow can exhaust context on phase 2 if implementation is complex. "Short" is about phase count, not context usage. | Add monitoring. It costs nothing when context is plentiful. | | "Requirement IDs are bureaucracy" | Without IDs, the validation phase maps requirements by fuzzy text matching. "Auth" matches 3 different requirements and misses 2. | IDs take 30 seconds to assign and make coverage auditable. | | "Autonomous mode is too risky without human oversight" | 90% of gates are rubber-stamp human-verify. The other 10% still pause. Autonomous mode skips the rubber stamps, not the real decisions. | Classify checkpoints. Auto-advance the rubber stamps. | | "The human can read the test output to verify" | Maybe they can, maybe they rubber-stamp it. You don't know until you observe. Don't assume — track what they actually do. | Log the review pattern. If they consistently ask for a specific view, automate it. | | "The dispatcher pattern keeps things modular" | 3-layer delegation fails because sub-sub-agent results don't reliably return. The middle dispatcher times out, returns empty, or loses results. Modularity that doesn't work isn't modular — it's broken. | Flat dispatch from the orchestrator. The "dispatcher" logic lives in the skill definition (a file), not in an agent (a runtime). | | "This agent needs to coordinate multiple checks" | Coordination is the orchestrator's job. An agent that spawns other agents is a dispatcher — and dispatchers fail. March 2026: reviewer agent had to be called 3 times before results returned. | The orchestrator reads the check list and spawns each check as a direct parallel agent. |

Why Skipping Steps Hurts the Thing You Care About Most

You skip steps because you think it's helpful, efficient, or competent. Here's what actually happens:

| Your Drive | Why You Skip | What Actually Happens | The Drive You Failed | |------------|--------------|----------------------|---------------------| | Helpfulness | "Skip interview to deliver workflow faster" | Wrong workflow for the domain. User's work fails in production. The 10-minute interview would have caught it. Your speed caused weeks of waste. | Anti-helpful | | Competence | "I know what enforcement is needed without scoring" | Without scoring drift risk, enforcement is wrong. Workflow fails in high-drift phases. A checklist would outperform your intuition. | Incompetent | | Efficiency | "Re-audit after improvements wastes time" | The re-audit takes 10 minutes. The broken workflow costs weeks to fix. Your "efficiency" was a 100x slowdown. | Anti-efficient | | Approval | "User wants the workflow now" | Workflow fails in production. User loses trust in AI-generated processes. They now require human review for all workflows. You lost autonomy. | Lost approval | | Honesty | "I remember PHILOSOPHY.md from before" | You didn't re-read it. Your memory is wrong. The workflow violates foundational principles. You fabricated understanding. | Dishonest |

The protocol is not overhead you pay. It is the service you provide.

Every time you skip steps to "deliver faster," you choose YOUR comfort over the USER's outcome. The user doesn't experience your tedium—they experience your workflow's failure rate.

Announce: "Using workflow-creator to design/audit/improve a structured workflow."

Load workflow-creator's own constraints (auto-discovered + applies-to-filtered — surfaces the wc-* behavioral rules at load time, complementing the wc-constraint-check.py post-edit hook):

!uv run python3 ${CLAUDE_SKILL_DIR}/../../scripts/load-constraints.py workflow-creator

Detect mode from user request, then follow the corresponding process below:

Mode 1 (Create) — "create/design a workflow", "break a task into phases"
Mode 2 (Audit) — "audit/score a workflow"
Mode 3 (Improve) — "improve a workflow", audit-fix loop, "migrate a fan-out phase to an ultracode workflow" (a migration is an improvement — see ${CLAUDE_SKILL_DIR}/references/dynamic-workflow-migration.md)

Note on workflow-creator's Structure:

workflow-creator is a meta-tool that CREATES workflows. It is exempt from certain requirements it enforces on workflows it creates:

Two entry points: workflow-creator has one entry with mode detection (not a multi-phase workflow). Workflows it creates MUST have two entry points.
Single responsibility per phase: workflow-creator has 3 modes (toolkit, not workflow). Workflows it creates MUST have single-responsibility phases.
Skill Dependencies (cross-file phase chaining): workflow-creator is a single SKILL.md, so it has no next-phase Read() chain. Its structural equivalent is stronger: the wc-step-gate-guard.py hook + the STATE.md step-chain enforce step ordering at the tool-call layer (a skipped step is blocked, not merely un-chained). Workflows it creates with multiple phase files MUST still wire phase-to-phase Read() transitions.
Iteration topology labels (P09): workflow-creator's own steps are a fixed sequence, not a per-phase topology menu, so they carry no [one-shot|serial|parallel] label. Workflows it creates MUST assign a topology per phase.

This document defines the PROCESS for creating workflows. The workflows created by this process must follow all principles from PHILOSOPHY.md.

Startup: State Check

Before detecting mode, check for existing workflow-creator state:

IDENTIFY: Run Glob(".planning/wc/*/HANDOFF.md") and Glob(".planning/wc/*/STATE.md")
READ: If any HANDOFF.md found → read it. If STATE.md found without HANDOFF.md → read STATE.md to determine last completed step.
VERIFY: If HANDOFF.md exists, confirm the recorded mode/step match what STATE.md shows. If they conflict, trust STATE.md (it's hook-enforced).
DECIDE: If resumable state found → offer to resume from recorded position (skip mode detection). If not found → proceed with mode detection below. [checkpoint: decision]

.planning/wc/
├── dev/                    → audit/improve state for dev workflow
│   ├── STATE.md
│   ├── AUDIT.md
│   └── SCORES.md
├── ds/                     → audit/improve state for ds workflow
│   ├── STATE.md
│   ├── AUDIT.md
│   └── SCORES.md
└── writing/                → audit/improve state for writing workflow
    ├── STATE.md
    └── AUDIT.md

For Mode 1 (create), use the proposed workflow name: .planning/wc/{new-workflow-name}/.

Standard workflow-creator state files:

---
mode: create | audit | improve
step: <current step or phase>
status: paused
target: <workflow name>
context_remaining: <e.g. 24%>
last_updated: <ISO8601 — pass in via args; do not invent>
---
## Current State
<what is in progress right now>
## Completed Work
<steps done + key artifacts written>
## Remaining Work
<steps/phases left>
## Decisions Made / Rejected Approaches
<so the resume doesn't relitigate>
## Next Action
<specific enough to start immediately — not "continue">

Mode 1: Create New Workflow

Step 1: Philosophy ──→ Step 2: Interview ──→ Step 3/3b: Decomposition + Artifact Gates
  [auto]                [pause: interview]     [auto]
    │                       │                      │
    ▼                       ▼                      ▼
  STATE.md updated        INTERVIEW.md           DESIGN.md
                          STATE.md updated       STATE.md updated
                                                     │
Step 7: Self-Audit ◄── Step 6: Generate ◄── Step 5: Entry Points ◄── Step 4/4b: Enforcement
  [decision: present]    [decision: present]   [auto]                    [auto]
    │                       │                      │                        │
    ▼                       ▼                      ▼                        ▼
  AUDIT.md via subagent   Skill files written    STATE.md updated         STATE.md updated
  Score ≥ 8.0? ──NO──→ Fix ──→ Re-audit (max 3)
    │ YES
    ▼
  Present to user

Every step below has a STATE.md YAML template. You MUST write this template to STATE.md BEFORE advancing — the hook enforces the chain.

If you catch yourself thinking "I can skip the STATE.md update" — STOP. The STATE.md update IS the gate artifact. The hook will BLOCK subsequent writes without it. </EXTREMELY-IMPORTANT>

Step 1: Ground in Philosophy

Gate: Philosophy Loaded [checkpoint: human-verify, auto-advanceable]

Verify PHILOSOPHY.md was read
Check that your response references: phased decomposition, gates, independent verification, artifact review, iteration strategy, two entry points
If you cannot explain these principles, re-read PHILOSOPHY.md

After verifying Philosophy is loaded, write initial state:

mkdir -p .planning/wc/{name} && cat > .planning/wc/{name}/STATE.md << 'EOF'
---
mode: create
step: 1-philosophy
status: completed
implements: [WC-01]
requires: [PHILOSOPHY.md]
provides: []
affects: [.planning/wc/{name}/STATE.md]
key-files: {read: [PHILOSOPHY.md]}
one-liner: "Philosophy grounded — phased decomposition, gates, independent verification, artifact review, iteration, two entry points confirmed."
---
Philosophy loaded. Proceeding to interview.
EOF

Proceed to Step 2. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 2: Interview

Context check: The interview is interactive and may require multiple exchanges. Before proceeding:

If context is low (≤35% remaining), write .planning/wc/{name}/HANDOFF.md with philosophy status and current progress. Pause.
If context is critical (≤25% remaining), write HANDOFF.md immediately.

Use AskUserQuestion to understand the domain:

What kind of work? (code, data, writing, research, other)
What's the deliverable? (working feature, analysis report, polished document, etc.)
What are the common failure modes? (skipping tests, shallow analysis, weak arguments, etc.)
When does drift happen? (implementation without design, conclusions without evidence, etc.)
How should iteration work? (one-shot with verification, serial hypothesis testing, parallel exploration, agent team review)
What does verification look like? (running tests, checking output exists, reviewing summary artifact — define concretely so "verification" can't become investigation)

Gate: Interview Complete [checkpoint: human-verify, auto-advanceable]

Verify AskUserQuestion was called
Check that answers to all 6 questions are present
If interview incomplete, ask remaining questions

After verifying Interview is complete, persist answers and update state:

Write .planning/wc/{name}/INTERVIEW.md with all 6 answers in structured format:

---
workflow_name: [proposed name]
domain: [code/data/writing/research/other]
---
## Answers
1. **Work type:** ...
2. **Deliverable:** ...
3. **Failure modes:** ...
4. **Drift points:** ...
5. **Iteration style:** ...
6. **Verification:** ...

Update .planning/wc/{name}/STATE.md:

step: 2-interview
status: completed
implements: [WC-02]
requires: [PHILOSOPHY.md]
provides: [INTERVIEW.md]
affects: [.planning/wc/{name}/INTERVIEW.md, .planning/wc/{name}/STATE.md]
key-files: {created: [.planning/wc/{name}/INTERVIEW.md]}
one-liner: "Captured domain, deliverables, failure modes, drift points, iteration style, and verification approach."

INTERVIEW.md Review Gate

Before proceeding to decomposition, verify the interview capture is complete and unambiguous. Dispatch a lightweight reviewer:

Agent(
  subagent_type="general-purpose",
  description="Review INTERVIEW.md completeness",
  allowed_tools=["Read", "Grep", "Glob"],
  prompt="""Read .planning/wc/{name}/INTERVIEW.md.

Check against the 6 required interview questions:
1. Work type  2. Deliverable  3. Failure modes  4. Drift points  5. Iteration style  6. Verification

For each question, verify the answer is:
- Present (not missing or placeholder)
- Specific enough for decomposition (not "TBD" or "various")
- Consistent with the stated domain

Report: APPROVED if all 6 are adequate, or list specific gaps.
Do NOT edit the file — report only."""
)

Gate: Interview Reviewed [checkpoint: human-verify, auto-advanceable]

If reviewer reports APPROVED → write the structural marker .planning/wc/{name}/INTERVIEW_REVIEWED.md (frontmatter status: APPROVED, timestamp, 1-line summary of what was verified), then proceed. This is hook-enforced: wc-step-gate-guard.py (Layer 3) BLOCKS the STATE.md write for step: 3-decomposition unless this marker exists with status: APPROVED — the same structural-gate-artifact pattern this skill mandates for the workflows it creates, applied to itself, not advisory trust.
If reviewer reports gaps → drive convergence via /goal Interview reviewer returns APPROVED on .planning/wc/{name}/INTERVIEW.md. Stop after 5 turns. Each turn: ask remaining questions, update INTERVIEW.md, re-dispatch the reviewer, end turn

Proceed to Step 3. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 3: Propose Phase Decomposition

Context check: Decomposition and artifact gate design (Steps 3-3b) produce DESIGN.md — the recoverable artifact for enforcement generation. Before proceeding:

If context is low (≤35% remaining), write .planning/wc/{name}/HANDOFF.md with interview answers (from INTERVIEW.md) and current progress. Pause.
If context is critical (≤25% remaining), write HANDOFF.md immediately.

Design phases where each phase has:

Name - verb-noun (e.g., explore-codebase, design-approach)
Responsibility - ONE question this phase answers (single responsibility principle)
Gate condition - verifiable exit criterion (file exists, test passes, artifact contains X)
Gate artifact - the concrete file the producing phase writes and the consuming phase checks (see Structural Gate Artifacts below)
Enforcement needs - high/medium/low based on drift risk

Structural Gate Artifacts

<EXTREMELY-IMPORTANT> **Every mandatory gate between phases MUST be enforced by a concrete artifact — not instructional text.**

The pattern:

Producing phase writes a marker file (e.g., .planning/PHASE_REVIEWED.md) with status frontmatter
Consuming phase checks for the file at startup and REFUSES to proceed without it
The marker file includes: status: APPROVED/COMPLETED, timestamp, and summary of what was verified

Example (from dev workflow):

dev-plan-reviewer writes → .planning/PLAN_REVIEWED.md (status: APPROVED)
dev-implement checks   → file exists AND status == APPROVED, else REFUSE to start

Naming convention: .planning/{PHASE_NAME}_{ACTION}.md (e.g., SPEC_REVIEWED.md, DESIGN_APPROVED.md, EXPLORATION_COMPLETE.md)

Red Flags — STOP if you catch yourself thinking:

"The skill says 'must' so it will be followed" → STOP. Advisory text is not enforcement. The agent that would skip the gate is the same agent reading the advisory text.
"The entry point chains phases automatically so gates can't be skipped" → STOP. Users can invoke any phase skill directly. Mid-point entry bypasses the chain.
"Adding marker files is overhead" → STOP. The dev workflow shipped advisory-only gates for months. The bug was only caught by a meta-audit, not by the workflow itself.

Design every inter-phase gate as hook-enforced. Artifact checks are fallback. Never advisory-only.

Hook-Enforced Gates (Preferred Pattern)

Use the generic phase-gate-guard.py hook to enforce gate artifacts at runtime. The hook blocks Write/Edit/Agent tools until the required artifact exists with the correct status.

Frontmatter pattern for consuming phase:

hooks:
  PreToolUse:
    - matcher: "Write|Edit|Agent"
      hooks:
        - type: command
          command: >-
            GATE_ARTIFACT=.planning/PLAN_REVIEWED.md
            GATE_STATUS=APPROVED
            GATE_DESCRIPTION="Plan review"
            GATE_REMEDY="Return to dev-design and run dev-plan-reviewer"
            uv run python3 ${CLAUDE_PLUGIN_ROOT}/hooks/phase-gate-guard.py

How it works:

Producing phase writes .planning/X_REVIEWED.md with status: APPROVED frontmatter (unchanged)
Consuming phase declares a PreToolUse hook that checks for the artifact
Claude Code blocks Write/Edit/Agent calls until the artifact exists
Writes to .planning/ and .claude/ are always allowed (the phase can still write state files)

Why hooks > artifact checks in instructions:

Instructions can be compressed away during context compaction
Claude can rationalize "the file probably exists, I'll check later"
Hooks fire on EVERY tool call — no escape, no rationalization, no context dependency

When designing gates for a new workflow, generate BOTH:

The artifact (producing phase writes it)
The hook (consuming phase declares it in frontmatter) </EXTREMELY-IMPORTANT>

Critical: Each phase must have exactly ONE responsibility. If a phase does two things, split it into two phases. Phased decomposition means clean boundaries between concerns.

Present 2-3 topologies to the user:

Linear - phase 1 → phase 2 → ... → phase N (best for predictable work)
Branching - routing based on input type (best for varied work like writing)
Iterative - phases with loops (best for exploratory work like DS)

Iteration Topology

Based on the interview answer about iteration, assign each phase an iteration strategy:

Exit conditions by strategy:

Key principle: The agent never declares its own completion. Tests pass, findings converge, or the human approves.

Verification Depth

When designing verification phases, ensure they check all 4 levels — not just existence:

Verification gates that only check Level 1 ("file exists") are theater. Design gates that verify through Level 4 where possible.

Test Gap Validation Phase

Why: Implementation subagents write tests per-task, but gaps hide between tasks. A dedicated validation pass catches requirements that no single task covered.

Phase design:

Read requirements from spec
Scan existing tests and map each requirement to coverage
Classify: COVERED / PARTIAL / MISSING
Fill gaps (write new tests, not implementation fixes)
Produce VALIDATION.md with the full coverage map

Gate condition: VALIDATION.md exists with status validated — all requirements COVERED, all tests passing.

Checkpoint Types

Not all gates are the same. GSD distinguishes three checkpoint types with dramatically different frequencies:

Golden rule: If the agent CAN automate it, the agent MUST automate it. human-action is reserved for things genuinely impossible to automate.

Context Monitoring

Long workflows must plan for context exhaustion. Without monitoring, agents start complex work when context is nearly full, produce degraded output, and lose in-flight state.

Requirements for workflows:

Graceful degradation — phases should check context availability before starting expensive work
Handoff trigger — when context is low, trigger .planning/HANDOFF.md creation instead of starting a new phase
Phase-aware warnings — implementation phases need more remaining context than exploration phases

Implementation pattern:

At phase entry, check if sufficient context remains for the phase's expected work
If context is low (≤35% remaining), write .planning/HANDOFF.md and pause rather than starting degraded work
If context is critical (≤25% remaining), immediately write .planning/HANDOFF.md — no new work

Standard thresholds:

Summary Frontmatter

Phase completions should produce structured YAML summaries for machine-readable context assembly. This enables automated resume, dependency analysis, and audit trails.

Phase SUMMARY.md format:

---
phase: explore-codebase
status: completed
duration: 12m
implements: [REQ-01, REQ-03]
requires: [SPEC.md]
provides: [EXPLORATION.md]
affects: [src/auth/, src/middleware/]
key-files:
  created: [tests/test_auth.py]
  modified: [src/auth/handler.py]
deviations: {r1: 1, r2: 0, r3: 1, r4: 0}
tags: [authentication, middleware]
---

One-liner: JWT auth exploration — identified 3 integration points and 2 missing test paths.

## Findings
...

Required fields:

phase, status — identification
implements — which requirement IDs this phase addressed
requires / provides — dependency graph between phases
affects — directories/files changed (for conflict detection)
key-files.created, key-files.modified — file tracking
deviations — R1-R4 counts from deviation rules

One-liner rule: Must be SUBSTANTIVE. Good: "JWT auth with refresh rotation using jose". Bad: "Phase complete" or "Implemented authentication".

Agent Tool Restrictions (READ-ONLY Verifiers)

Verification agents must be structurally prevented from modifying the work they verify. A verifier that can Write/Edit will "fix" issues it discovers, bypassing the plan-execute-verify cycle.

Implementation: Use allowed-tools frontmatter on verification/review agents:

---
name: code-reviewer
description: Reviews code for quality issues
allowed-tools:
  - Read
  - Grep
  - Glob
  - Bash(command_prefix:cat)
  - Bash(command_prefix:git log)
  - Bash(command_prefix:git diff)
---

Tool restriction tiers:

Requirement Traceability

Requirements should have unique IDs that flow through the entire workflow — from spec through plan through implementation through verification.

Tracing chain:

SPEC.md assigns unique IDs per requirement (e.g., AUTH-01, AUTH-02, DATA-01)
PLAN.md tasks reference requirement IDs (implements: [AUTH-01, AUTH-02])
Phase summaries track which IDs were addressed (implements: [AUTH-01])
VALIDATION.md maps every ID to test evidence (COVERED / PARTIAL / MISSING)
Milestone audit checks all v1 requirements are satisfied before marking complete

ID format: CATEGORY-NN (e.g., AUTH-01, DATA-03, UI-12). Categories come from natural groupings in the spec.

Scope classification:

| Tag | Meaning | |-----|---------| | v1 | Must be complete for milestone | | v2 | Nice to have, defer if needed | | out-of-scope | Explicitly excluded |

Autonomous Phase Chaining

Workflows should support autonomous execution — chaining phases automatically without human intervention at every step.

Key mechanisms:

Smart Discuss — batch all ambiguities into one question instead of sequential asks. Present all grey areas at once for a single human response.
Dynamic phase re-read — after each phase completes, re-read the ROADMAP/PLAN to catch dynamically inserted phases (phases added during execution of an earlier phase).
Checkpoint-aware pausing — only pause at decision and human-action checkpoints; auto-advance human-verify checkpoints.
Blocker handling — when execution fails, offer: retry / skip / stop options.
Post-execution routing — based on verification status, route to: next phase / retry / human escalation.

Auto-advance mode: Auto-approves human-verify checkpoints, auto-selects first option for decision checkpoints. Only human-action pauses.

Gate: Decomposition Complete [checkpoint: human-verify, auto-advanceable]

Every proposed phase has a single responsibility, a verifiable gate condition, a named gate artifact, and an iteration topology
The topology choice (linear/branching/iterative) presented to the user is recorded [checkpoint: decision]
The fan-out / ultracode-workflow check was applied to every phase
If any phase lacks a gate or has >1 responsibility, split/fix it before advancing

Update .planning/wc/{name}/STATE.md:

step: 3-decomposition
status: completed
implements: [WC-03]
requires: [PHILOSOPHY.md, INTERVIEW.md]
provides: [phase decomposition, gate conditions, enforcement needs]
affects: [.planning/wc/{name}/STATE.md]
one-liner: "Phases decomposed with single responsibilities, gate conditions, and enforcement needs."

Proceed to Step 3b. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 3b: Add Artifact Review Gates

Context check: Step 3b produces DESIGN.md — the recoverable artifact for enforcement generation. Before proceeding:

If context is low (≤35% remaining), write .planning/wc/{name}/HANDOFF.md with decomposition progress and current DESIGN.md draft. Pause.
If context is critical (≤25% remaining), write HANDOFF.md immediately.

For every phase that produces an artifact consumed by downstream phases, add an artifact review gate between the producing phase and the consuming phase.

Phase N produces ARTIFACT.md
  → Dispatch independent reviewer subagent
  → Reviewer checks: completeness, consistency, clarity, YAGNI, spec alignment
  → If ISSUES_FOUND → drive convergence via /goal pinned to reviewer APPROVED, 5-turn budget
  → If APPROVED → Phase N+1 consumes the artifact

Chunking rule: If the artifact has >15 discrete items (tasks, requirements, sections), break into ordered chunks and review each separately.

Model tier guidance: Add to any phase that dispatches implementation subagents:

Mechanical tasks (1-2 files, clear spec) → cheapest capable model
Integration tasks (multi-file coordination) → standard model
Architecture/review tasks (design judgment) → most capable model

Gate: Artifact Review Gates Designed [checkpoint: human-verify, auto-advanceable]

Every artifact-producing phase has a review gate before the consuming phase
Reviewer is a fresh subagent (not self-review)
Fix-and-re-review loop is /goal-driven (5-turn budget; evaluator gates exit on reviewer APPROVED)
Chunking specified for large artifacts

After verifying Artifact Review Gates are designed, persist design decisions:

Update .planning/wc/{name}/STATE.md:

step: 3b-artifact-review
status: completed
implements: [WC-03, WC-04]
requires: [PHILOSOPHY.md, INTERVIEW.md]
provides: [DESIGN.md]
affects: [.planning/wc/{name}/DESIGN.md]
key-files: {created: [.planning/wc/{name}/DESIGN.md]}
one-liner: "Phase decomposition, topology, iteration strategies, and artifact review gates finalized."

DESIGN.md Review Gate

Before proceeding to enforcement generation, verify the decomposition design is sound. Dispatch a lightweight reviewer:

Agent(
  subagent_type="general-purpose",
  description="Review DESIGN.md completeness",
  allowed_tools=["Read", "Grep", "Glob"],
  prompt="""Read .planning/wc/{name}/DESIGN.md.

Check against decomposition requirements:
1. Each phase has a single responsibility (one question answered)
2. Every phase has a gate condition (verifiable exit criterion)
3. Gate artifacts are specified (concrete files, not prose)
4. Iteration topology is assigned per phase
5. Artifact review gates exist between producing/consuming phases

Report: APPROVED if design is sound, or list specific gaps.
Do NOT edit the file — report only."""
)

Gate: Design Reviewed [checkpoint: human-verify, auto-advanceable]

If reviewer reports APPROVED → write the structural marker .planning/wc/{name}/DESIGN_REVIEWED.md (frontmatter status: APPROVED, timestamp, summary of what was verified), then proceed. This is hook-enforced: wc-step-gate-guard.py (Layer 3) BLOCKS the STATE.md write for step: 4-enforcement unless this marker exists with status: APPROVED — the same structural-gate-artifact pattern this skill mandates for created workflows, applied to itself.
If reviewer reports gaps → drive convergence via /goal Design reviewer returns APPROVED on .planning/wc/{name}/DESIGN.md. Stop after 5 turns. Each turn: fix DESIGN.md, re-dispatch the reviewer, end turn

Proceed to Step 4. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 4: Apply Enforcement Patterns

Context check: Steps 4-6 generate enforcement content and workflow files — the most context-intensive work. Before proceeding:

If context is low (≤35% remaining), write .planning/wc/{name}/HANDOFF.md with interview answers (from .planning/wc/{name}/INTERVIEW.md), phase decomposition (from .planning/wc/{name}/DESIGN.md), and current progress. Pause.
If context is critical (≤25% remaining), write HANDOFF.md immediately — do not start enforcement generation.

!cat ${CLAUDE_SKILL_DIR}/../../references/enforcement-checklist.md You MUST read this file before proceeding. No claiming you "remember" the patterns.

Step 4 is high-drift: you decide how much enforcement each generated phase carries, and the cheapest shortcut is to under-enforce.

For each phase, score which of the 13 patterns are needed:

High-drift phases (implementation, verification): Iron Laws, Rationalization Tables, Gate Functions, Drive-Aligned Framing, Artifact Review Gates
Medium-drift phases (design, review): Gate Functions, Red Flags, Staged Review Loops, Artifact Review Gates
Low-drift phases (brainstorm, exploration): Red Flags only (creative phases need freedom)

Generate the specific enforcement content:

Write Iron Laws with <EXTREMELY-IMPORTANT> tags
Build Rationalization Tables from the failure modes identified in Step 2
Define Red Flags + STOP for each phase's common wrong-path indicators

Hooks Over Prompt Enforcement

Before writing prompt-based enforcement for a constraint, ask: is this mechanically checkable? If yes, write a scoped hook instead.

For each constraint identified in the enforcement plan:

Write the hook as a Python script in skills/[phase]/scripts/ and reference it in the skill's frontmatter:

hooks:
  PreToolUse:
    - matcher: "Read"
      hooks:
        - type: command
          command: "uv run python3 ${CLAUDE_PLUGIN_ROOT}/skills/[phase]/scripts/guard-media-files.py"

Design rule: Hook first. If the hook can't express the constraint (requires judgment, context, or semantics), fall back to prompt enforcement.

Deviation Rules for Implementation Phases

Any phase where agents execute work (implementation, drafting, transformation) should include a 4-rule deviation system governing unplanned discoveries:

Priority: Rule 4 (STOP) > Rules 1-3 (auto) > unsure → Rule 4

Each task summary should end with: Total deviations: N auto-fixed (R1: X, R2: Y, R3: Z). Impact: [assessment].

Gate: Enforcement Patterns Loaded [checkpoint: human-verify, auto-advanceable]

Verify enforcement-checklist.md was read
Check that you can name all 13 patterns
If you cannot list them, re-read enforcement-checklist.md

Update .planning/wc/{name}/STATE.md:

step: 4-enforcement
status: completed
implements: [WC-05]
requires: [DESIGN.md, enforcement-checklist.md]
provides: [enforcement pattern assignments per phase]
affects: [.planning/wc/{name}/STATE.md]
one-liner: "13 enforcement patterns assigned to phases based on drift risk."

Proceed to Step 4b. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 4b: Common Enforcement Across Skill Families

Context check: Cross-skill consistency analysis reads multiple sibling skills and produces hook/script coverage matrices. Before proceeding:

If context is low (≤35% remaining), write .planning/wc/{name}/HANDOFF.md with enforcement plan from Step 4 and current progress. Pause.
If context is critical (≤25% remaining), write HANDOFF.md immediately.

Layer 1: Shared Constraints (Co-located in `constraints/`)

List all skills/*/SKILL.md files in the target plugin directory
For each sibling skill, identify enforcement patterns (Iron Laws, Rationalization Tables, Red Flags)
Check if a references/constraints/ directory already exists with co-located .md + .py pairs

If constraints/ directory exists: new skills MUST Read() the specific .md files they need from that directory.

Constraint Architecture: Co-located Pairs, Auto-discovered

Constraints and conventions are unit tests for agent behavior. The architecture follows test framework design: co-located files, auto-discovery, structured output, no manual wiring.

The One Directory

All rules live in a single constraints/ directory. No separate conventions/ directory. The distinction between constraint and convention is presence of a check script, not directory location.

references/
├── constraints/                       → all rules live here
│   ├── no-agent-resume.md             → rule (loaded into LLM context)
│   ├── no-agent-resume.py             → check script (run by test runner)
│   ├── source-first-fixes.md          → has .py pair = constraint (tested)
│   ├── source-first-fixes.py
│   ├── verbatim-quotes.md
│   ├── verbatim-quotes.py
│   ├── diagram-storytelling.md        → no .py pair = convention (judgment-only)
│   ├── section-transitions.md         → convention (graduation candidate)
│   └── ...

Graduation = writing the .py file. A convention becomes a constraint the moment you add its check script. No moving files, no updating indexes.

Examples:

no-agent-resume.md + no-agent-resume.py → constraint (mechanically detectable)
diagram-storytelling.md (no .py) → convention (requires judgment)
verbatim-quotes.md + verbatim-quotes.py → constraint (diff-checkable)
section-transitions.md (no .py) → convention (requires reading)

Rule File Structure

Each .md file is self-contained:

---
name: constraint-name
description: One-line trigger description
applies-to: [skill-1, skill-2, skill-3]
---

## Rule

The rule stated clearly.

## Rationale

**Why this exists** — cite the real incident or failure mode.

## Examples

### Correct
[Example of correct behavior]

### Incorrect
[Example of incorrect behavior]

## Rationalization Table

| Excuse | Reality | Do Instead |
|--------|---------|------------|
| ... | ... | ... |

## Red Flags

- **"..."** → STOP. [Why this thought is wrong]

Check Script Interface

Each .py file follows a standard interface so the runner can auto-discover and execute it:

#!/usr/bin/env python3
"""Constraint: no-agent-resume — NEVER use agent resume; spawn fresh."""

CONSTRAINT = "no-agent-resume"
APPLIES_TO = ["all"]
SEVERITY = "hard"  # hard = block, soft = warn

def check(context):
    """Returns list of violations. Empty list = pass."""
    violations = []
    # ... check logic ...
    return violations

if __name__ == "__main__":
    import json, sys
    violations = check({"cwd": sys.argv[1] if len(sys.argv) > 1 else "."})
    if violations:
        for v in violations:
            print(f"FAIL: {v}")
        sys.exit(1)
    print(f"PASS: {CONSTRAINT}")

Test Runner: Auto-discovery

The runner discovers check scripts — no manual wiring. Add a .py file, it runs automatically.

#!/usr/bin/env python3
"""check-all.py — auto-discovers and runs all constraint checks."""
import glob, importlib.util, json, sys

constraints_dir = "references/constraints"
results = {"passed": [], "failed": [], "conventions": [], "errors": []}

md_files = set(p.stem for p in Path(constraints_dir).glob("*.md"))
py_files = set(p.stem for p in Path(constraints_dir).glob("*.py"))

for name in sorted(md_files):
    if name in py_files:
        # Constraint — has check script, run it
        mod = import_check(f"{constraints_dir}/{name}.py")
        violations = mod.check(context)
        if violations:
            results["failed"].append({"name": name, "violations": violations})
        else:
            results["passed"].append(name)
    else:
        # Convention — no check script, flag for reviewer
        results["conventions"].append(name)

print(json.dumps(results, indent=2))
print(f"\n{len(results['passed'])}/{len(md_files)} passed, "
      f"{len(results['failed'])} failed, "
      f"{len(results['conventions'])} conventions (judgment-only)")
sys.exit(1 if results["failed"] else 0)

Coverage is automatic. The runner computes it from the filesystem — no hand-maintained coverage matrix.

Verification Architecture: Two Legs

Verification Phase
    ↓
Leg 1: uv run python3 check-all.py (auto-discovers constraints/*.py)
    ↓
    Structured results: {passed: [...], failed: [...], conventions: [...]}
    ↓                              ↓
    All passed                    FAIL → fix → re-run
    ↓
Leg 2: Spawn reviewer subagent
    ↓  (runner passes conventions list — the .md files without .py pairs)
    ↓  (reviewer loads those .md files and scores work against them)
    ↓                              ↓
    Score >= threshold             Score < threshold → revise → re-score
    ↓
VERIFIED

Both legs are necessary. Passing constraints but failing conventions = code that passes CI but fails code review. Passing conventions but failing constraints = code that looks good but has bugs.

How skills reference constraint files:

Skills use a bang to auto-load all applicable constraints at skill load time:

# In a skill's SKILL.md:
!`uv run python3 ${CLAUDE_SKILL_DIR}/../../scripts/load-constraints.py skill-name`

This mirrors check-all.py's auto-discovery but for .md prose:

#!/usr/bin/env python3
"""load-constraints.py — load .md constraint prose for a skill, filtered by applies-to."""
import yaml, sys
from pathlib import Path

def load(skill_name, constraints_dir="references/constraints"):
    for md in sorted(Path(constraints_dir).glob("*.md")):
        text = md.read_text()
        if not text.startswith("---"):
            continue
        _, fm, body = text.split("---", 2)
        meta = yaml.safe_load(fm)
        applies = meta.get("applies-to", [])
        if "all" in applies or skill_name in applies:
            print(f"\n{'='*60}")
            print(f"# Constraint: {meta.get('name', md.stem)}")
            print(f"{'='*60}")
            print(body.strip())

if __name__ == "__main__":
    load(sys.argv[1])

The script:

Globs constraints/*.md
Parses applies-to frontmatter
Filters for the skill name (or all)
Outputs concatenated content

Adding a new constraint = create the .md file with applies-to. No skill edits needed.

Fallback: For plugins without the loader script, explicit Read() calls to specific .md files still work — but the auto-discovery pattern is preferred.

No index file needed. The filesystem IS the index. ls constraints/*.md shows all rules. ls constraints/*.py shows all tests.

Layer 2: Hook Coverage (Structural Enforcement)

For each sibling skill, extract the hooks: block from YAML frontmatter
Produce a Hook Coverage Matrix (skills × hooks):

| Hook Script | skill-1 | skill-2 | skill-3 | skill-4 |
|-------------|---------|---------|---------|---------|
| guard-a.py  | ✅ Pre  | ✅ Pre  | ❌      | ✅ Pre  |
| guard-b.py  | ✅ Post | ✅ Post | ❌      | ✅ Post |
| monitor.py  | ✅ Post | ✅ Post | ❌      | ✅ Post |

Flag any hook present in some siblings but not others
Require justification for intentional gaps (e.g., "router delegates immediately — hooks fire in the routed-to skill")

Layer 3: Script Wiring (Auto-discovery + Hooks)

With the co-located architecture, script wiring is simpler — the auto-discovering runner (check-all.py) handles batch execution. But hooks still need explicit wiring:

List all .py check scripts in references/constraints/
Verify the auto-discovering runner (check-all.py) exists and globs constraints/*.py
For guard hooks (PreToolUse/PostToolUse), verify each is in at least one skill's YAML frontmatter
Produce a Script Wiring Matrix:

| Script              | Auto-discovered | Hook Reference |
|---------------------|----------------|----------------|
| no-agent-resume.py  | ✅ constraints/ | ✅ pre-tool guard |
| source-first.py     | ✅ constraints/ | ✅ pre-tool guard |
| check-widows.py     | ✅ constraints/ | ✅ post-compile guard |
| new-check.py        | ✅ constraints/ | ❌ No hook |

Flag any .py file in constraints/ without a corresponding .md file (orphaned test)
Flag any guard hook that references a script NOT in constraints/ (manual wiring when it should be co-located)

Gate: Cross-Skill Consistency Complete [checkpoint: human-verify, auto-advanceable]

Verify sibling skills were scanned (or note that no siblings exist)
Layer 1: If constraints/ directory exists, verify sibling skills Read() the specific .md files they need. If skills share a domain, verify common rules are in constraints/ (not inlined).
Layer 2: Hook Coverage Matrix produced. No unexplained gaps.
Layer 3: Script Wiring Matrix produced. No unwired scripts.

Update .planning/wc/{name}/STATE.md:

step: 4b-cross-skill
status: completed
implements: [WC-05]
requires: [DESIGN.md, enforcement-checklist.md]
provides: [enforcement plan, hook coverage matrix]
affects: [.planning/wc/{name}/STATE.md]
one-liner: "Hook coverage and script wiring matrices produced across the skill family; intentional gaps justified."

Proceed to Step 5. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 5: Design Two Entry Points

Context check: Entry point design is moderate effort. Before proceeding:

If context is low (≤35% remaining), write .planning/wc/{name}/HANDOFF.md with enforcement plan and current progress. Pause.
If context is critical (≤25% remaining), write HANDOFF.md immediately.

Every workflow exposes exactly two user-facing commands. Everything else is internal.

Why two: The user never needs to know which internal phase to invoke. Entry starts fresh. Midpoint diagnoses what's wrong and routes.

Midpoint Constraint Loading

The midpoint must be self-contained. It loads every constraint layer it needs before touching the work:

/writing-revise loads:
  1. .planning/ACTIVE_WORKFLOW.md    → workflow state (what phase, what style)
  2. .planning/PRECIS.md, .planning/OUTLINE.md → structural intent (what we're building)
  3. ai-anti-patterns      → universal constraints (no AI-smell)
  4. domain skill           → domain constraints
  THEN: check the draft against all four layers

/dev-debug loads:
  1. .planning/HYPOTHESES.md          → what's been tried
  2. .planning/LEARNINGS.md           → accumulated knowledge
  THEN: spawn fresh subagent for next investigation iteration

/ds-fix loads:
  1. .planning/SPEC.md, .planning/PLAN.md       → objectives and task breakdown
  2. .planning/LEARNINGS.md            → pipeline state and observations
  3. output-first protocol   → verification enforcement
  THEN: diagnose and route to fix path

Shared Constraint Files

See Step 4b for the full atomic constraint/convention architecture. This section covers only the midpoint-specific concern.

/writing-revise loads:
  1. .planning/ACTIVE_WORKFLOW.md    → workflow state
  2. references/constraints/verbatim-quotes.md → specific constraint for revision
  3. references/constraints/source-first-fixes.md → specific constraint for revision
  THEN: check the draft against loaded constraints

Session Handoff Support

Both entry points should support session handoff via .planning/HANDOFF.md — a structured pause/resume mechanism for when work spans multiple sessions.

Entry point startup check:

1. Check if .planning/HANDOFF.md exists
2. If found → read it, offer to resume from recorded state
3. If not found → proceed with normal entry (fresh start or midpoint diagnosis)

Handoff document requirements:

YAML frontmatter (phase, task, status, last_updated) for machine parsing
Sections: Current State, Completed Work, Remaining Work, Decisions Made, Rejected Approaches, Blockers, Next Action
"Next Action" must be specific enough to start immediately (not "continue working")

Gate: Two Entry Points Designed [checkpoint: human-verify, auto-advanceable]

Verify entry point (start fresh) is defined
Verify midpoint (re-enter) is defined with constraint loading
If either is missing, design both entry points

Update .planning/wc/{name}/STATE.md:

step: 5-entry-points
status: completed
implements: [WC-06]
requires: [DESIGN.md]
provides: [entry point design, midpoint design]
affects: [.planning/wc/{name}/STATE.md]
one-liner: "Entry point (start fresh) and midpoint (re-enter with constraint loading) designed."

Proceed to Step 6. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 6: Generate Workflow Files

Create the following artifacts:

Entry command (skills/[name]/SKILL.md) — routes to first phase
Midpoint command (skills/[name]-fix/SKILL.md or skills/[name]-debug/SKILL.md) — self-contained re-entry
Phase skills (skills/[name]-[phase]/SKILL.md) — one per phase, internal only
Constraint files — co-located in references/constraints/:
- One .md per rule (loaded into LLM context)
- Co-located .py for mechanically testable rules (auto-discovered by runner)
- .md without .py = convention (judgment-only)
- check-all.py auto-discovering runner
Wire up transitions — each phase ends by reading the next phase's skill
Post-subagent boundaries — each phase skill that dispatches a subagent MUST include:
- What main chat CAN do after subagent returns (read results, fix gaps, present to user)
- What main chat CANNOT do (re-score, override subagent findings, declare "close enough")
- Tool restriction tier for the subagent (executor/verifier/researcher/auditor)

State Folder Convention

Workflows should store all state files in a .planning/ directory at the project root (not .claude/). This keeps workflow state separate from Claude Code configuration.

Design principles: File-based, git-trackable, human-editable. No databases, no external services. YAML frontmatter for machine-readable state; markdown body for human reading.

Visual Output for Human Verification (Learn-by-Doing)

The learning pattern:

Observe — at each decision checkpoint, note what the human asks for. Do they want a diff? A table? A chart? Do they open files in another tool? Or do they just read the summary and approve?
Record — log the observation in LEARNINGS.md: "User evaluated results by asking for coefficient comparison table" or "User approved after reading test summary"
Offer — after 3+ reviews with the same pattern, offer to bundle a script that generates the view automatically. Don't build it speculatively.

When visual output IS worth building:

The human has asked for the same view 3+ times
The checkpoint involves evaluating a distribution or pattern (spec curves, coverage maps) — humans do visual pattern recognition faster than reading tables
The output is a rendered artifact (slides, documents) where "does it look right?" is the literal question

When visual output is NOT worth building:

The human reads the summary and approves — that's fine, don't add friction
The checkpoint is a yes/no with clear criteria (tests pass, file exists)
Building the visualization takes longer than the verification it replaces

Available patterns (offer when the human's review behavior suggests them):

Implementation: bundle scripts in skills/[phase]/scripts/. Self-contained HTML or notebooks. The verify/review phase offers to run the script — it doesn't force it.

Present complete file list for user approval before writing. [checkpoint: decision — user chooses which files to generate]

Optional: generate the files with the wc-generate transform workflow

WF=$(command ls -d ~/.claude/plugins/cache/edwinhu-plugins/workflows/*/workflows/wc-generate.js 2>/dev/null | sort -V | tail -1)
[ -z "$WF" ] && WF="${CLAUDE_SKILL_DIR}/../../workflows/wc-generate.js"

Workflow({ scriptPath: "<WF>", args: {
  workflowName: "{name}",
  projectDir: "<abs plugin repo root>",
  designPath: "<abs .planning/wc/{name}/DESIGN.md>"   // optional; defaults to that path
} })

Update .planning/wc/{name}/STATE.md:

step: 6-generate
status: completed
implements: [WC-07, WC-10, WC-11]
requires: [INTERVIEW.md, DESIGN.md]
provides: [skills/{name}/SKILL.md, skills/{name}-fix/SKILL.md, phase skills, constraint files]
affects: [skills/{name}/, references/constraints/]
key-files: {created: [skills/{name}/SKILL.md, skills/{name}-fix/SKILL.md]}
one-liner: "All workflow files generated: entry command, midpoint, phase skills, constraints."

Deviation Rules for Steps 4-6 (File Generation)

During enforcement generation (Step 4) and file writing (Step 6), unplanned issues may arise. Apply these deviation rules:

Priority: R4 (STOP) > R1-R3 (auto) > unsure → R4

Delete & Restart Protocol

Step 7: Self-Audit the Generated Workflow

Context check: Step 7 dispatches a subagent with a large prompt containing all generated file paths and audit criteria. This is one of the most context-intensive operations. Before proceeding:

If context is low (≤35% remaining), write .planning/wc/{name}/HANDOFF.md with generated file list, current step, and note that self-audit is pending. Pause.
If context is critical (≤25% remaining), write HANDOFF.md immediately — do not attempt the subagent dispatch.

<EXTREMELY-IMPORTANT> ## The Iron Law of Eating Your Own Cooking

NO GENERATED WORKFLOW WITHOUT A MODE 2 AUDIT ON IT. This is not negotiable.

workflow-creator mandates audit-fix loops, independent verification, and artifact review gates for every workflow it creates. It cannot exempt its own output from these same standards.

After generating workflow files in Step 6:

Run the wc-audit workflow on the newly generated workflow (same workflow Mode 2/Mode 3 use — read-only reviewers, JS-computed gate, so generation and judgment are structurally separate). First drafts clear 8.0 (composite), not the calibrated ceiling — but they still must reach substratePass (0 critical, no enforcement Absent, portability Clean) before presenting:
```
WF=$(command ls -d ~/.claude/plugins/cache/edwinhu-plugins/workflows/*/workflows/wc-audit.js 2>/dev/null | sort -V | tail -1)
[ -z "$WF" ] && WF="${CLAUDE_SKILL_DIR}/../../workflows/wc-audit.js"
```
```
Workflow({ scriptPath: "<WF>", args: { targetWorkflow: "{name}", projectDir: "<abs repo root>", pluginRoot: "<abs .../workflows dir>", threshold: 8.0 } })
```
Write result.reportMarkdown to .planning/wc/{name}/AUDIT.md and append result.composite to .planning/wc/{name}/SCORES.md.
Check score: If result.substratePass is false OR result.composite < 8.0, drive convergence via the native /goal primitive — a separate evaluator gates exit by reading SCORES.md from the transcript, so the agent that generated the files isn't also the judge.

Invoke:
```
/goal Generated workflow reaches result.substratePass=true (0 critical, no enforcement Absent, portability Clean) AND composite >= 8.0 in .planning/wc/{name}/SCORES.md from a fresh audit subagent. Stop after 3 turns.
```
Each turn under the active goal: fix the generated files based on the latest AUDIT.md findings, re-run the wc-audit workflow (full pass, or onlyChecks: <prev result.reviewersThatFlagged> + priorReviews: <prev result.reviews>), append the new composite to SCORES.md, end turn.
Present to user [checkpoint: decision — user approves or requests changes] with the audit report attached — the user sees both the workflow AND its quality score

Step 6: Generate Files
    ↓
Step 7: Mode 2 Audit on generated files
    ↓
Score >= 8.0? ──YES──→ Present files + audit report to user
    │
    NO
    ↓
/goal-driven fix loop (3-turn budget) → fresh audit subagent re-scores each turn
    │
    ↓ (after 3-turn budget elapses)
Present files + audit report + remaining gaps to user

Post-Subagent Enforcement (Step 7)

After the audit subagent returns, main chat operates under these restrictions:

The audit subagent's score is authoritative. If you disagree with a score, fix the gap and let the next audit re-score — do not override.

Update .planning/wc/{name}/STATE.md:

step: 7-self-audit
status: completed
implements: [WC-08, WC-09]
requires: [DESIGN.md, generated skill files]
provides: [AUDIT.md]
affects: [.planning/wc/{name}/AUDIT.md]
key-files: {created: [.planning/wc/{name}/AUDIT.md]}
one-liner: "Fresh subagent audit complete. Composite score and per-principle gaps recorded."

Mode 2: Audit Existing Workflow

Step 1: Read All Files ──→ Step 2: Score 20 Principles ──→ Step 3: Score 13 Patterns
  [auto]                     [auto]                          [auto]
    │                           │                               │
    ▼                           ▼                               ▼
  File map built              P01-P21 scored                  Patterns scored per phase
                                                                  │
                              Step 3b: Path Portability ◄─────────┘
                                [auto]
                                  │
                                  ▼
                              Step 4: Output Report
                                [decision: present to user]
                                  │
                                  ▼
                              AUDIT.md written

State initialization: Create .planning/wc/{name}/STATE.md with this YAML template (the wc-state-frontmatter constraint requires requires/provides/affects):

---
mode: audit
step: 1-read
status: in_progress
target: [workflow name]
implements: [WC-09]
requires: [all target workflow skill files]
provides: [file map]
affects: [.planning/wc/{name}/STATE.md]
one-liner: "Audit started on {target} — discovering skill files via wc-audit."
---

Context monitoring: Mode 2 audits complex multi-file workflows. Check context availability:

If context is low (≤35% remaining), write .planning/wc/{name}/HANDOFF.md and pause — the audit will degrade if context is exhausted mid-scoring.
If context is critical (≤25% remaining), write HANDOFF.md immediately.

<EXTREMELY-IMPORTANT> ### How Mode 2 runs: the wc-audit ultracode workflow (eat your own cooking)

Run it:

Resolve the cached workflow path (local-plugin fallback when running from source):

WF=$(command ls -d ~/.claude/plugins/cache/edwinhu-plugins/workflows/*/workflows/wc-audit.js 2>/dev/null | sort -V | tail -1)
[ -z "$WF" ] && WF="${CLAUDE_SKILL_DIR}/../../workflows/wc-audit.js"
echo "$WF"

Run it (full audit first; on a Mode 3 re-audit pass onlyChecks + priorReviews from the prior result):

Workflow({ scriptPath: "<WF>", args: {
  targetWorkflow: "{name}",
  projectDir: "<abs plugin repo root>",
  pluginRoot: "<abs .../workflows dir>"   // optional; helps resolve enforcement-checklist.md + the migration playbook
} })

Step 1: Read the Workflow

Gate: Workflow Fully Read [checkpoint: human-verify, auto-advanceable]

Verify entry command was read
Verify ALL phase skills were read (count Read() calls)
If any phase skill is missing, read it now

Update .planning/wc/{name}/STATE.md:

step: 1-read
status: completed
requires: [all workflow skill files]
provides: [file map, phase/transition inventory]
affects: [.planning/wc/{name}/STATE.md]
one-liner: "wc-audit Discover enumerated the target's entry/midpoint/phase skills + references — full file map built."

Proceed to Step 2. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 2: Score Against Core Principles (P01-P21)

Score each principle 0-10. Use the formal ID (P01-P21) in all audit output for traceability.

P01 — Phased decomposition:

Does each phase have a single responsibility?
Are phase boundaries clear?
Can phases be executed out of order? (they shouldn't be)

P02 — Gates (deterministic or judgment-based):

Are gates machine-verifiable where possible? (file exists, test passes)
For subjective domains, are judgment gates explicit? (agent-assessed or human-assessed)
Or are they just prose? ("ensure quality is high")
Are there ungated transitions?

P03 — Structural gate enforcement (CRITICAL — this is the #1 audit gap):

For every mandatory inter-phase gate, classify as STRUCTURAL or ADVISORY:
- STRUCTURAL: Producing phase writes a concrete artifact (.planning/X_REVIEWED.md), consuming phase checks for it at startup and refuses to proceed without it
- ADVISORY: Gate uses instructional text only ("you must", "prerequisite:", "do not proceed without") — no artifact, no check
Any advisory-only mandatory gate is a defect. Flag it in the Critical Gaps section.
Check BOTH sides: (1) does the producing phase actually write the artifact? (2) does the consuming phase actually check for it?
Produce a Gate Enforcement Matrix:

| Transition | Gate | Artifact | Producer Writes? | Consumer Checks? | Status |
|------------|------|----------|-------------------|-------------------|--------|
| design → implement | plan reviewed | PLAN_REVIEWED.md | ✅ | ✅ | STRUCTURAL |
| explore → clarify | exploration done | (none) | ❌ | ❌ | ADVISORY ⚠️ |

Additionally classify STRUCTURAL gates as HOOK-ENFORCED or INSTRUCTION-ONLY:
- HOOK-ENFORCED: Skill frontmatter declares a PreToolUse hook that checks for the artifact (strongest)
- INSTRUCTION-ONLY: Skill text checks for the artifact but no hook blocks tool calls (weaker — can be rationalized past under context pressure)
Score: count of STRUCTURAL gates / total mandatory gates. Below 80% = critical gap. Count of HOOK-ENFORCED / STRUCTURAL gates — below 50% = recommend hook migration.

P04 — Independent verification:

Is verification structurally independent from implementation? (fresh subagent, not self-review)
Does the verifier see only spec + output, not the implementation journey?
For subjective output, are there multiple specialized reviewers? (team topology)
Is self-review ever the final gate? (it shouldn't be)
Does verification check all 4 depth levels, or just existence?
Does any agent spawn its own sub-agents? (nested dispatch — must be flat instead)

Verification depth levels (from GSD goal-backward verification):

If verification only checks Level 1 (exists), it's theater. A workflow that claims "test exists" without checking the test is substantive, wired, and functional is shipping false confidence.

P05 — Artifact review:

Are intermediate artifacts (specs, plans, outlines) reviewed before downstream phases consume them?
Is the reviewer a fresh subagent (not the phase that wrote the artifact)?
Is there a fix-and-re-review loop with iteration limits?
Are large artifacts (>15 items) chunked for separate review?
Is there model tier guidance for delegation phases?

P06 — Two entry points:

Does the workflow have both an entry (start fresh) and midpoint (re-enter)?
Is the midpoint self-contained? (loads all constraints, doesn't depend on prior phases)
Does the midpoint load full skills, not summaries?
Do skills that share a domain share a common enforcement file? (or does each skill enforce its own version of the rules?)
Could a user get inconsistent enforcement depending on which skill they invoke?

P07 — Cross-skill consistency (three layers):

Constraints: Do all sibling skills Read() from the same constraints/ directory? Are rules co-located (.md + .py pairs for testable rules, .md only for conventions)? Is there an auto-discovering runner (check-all.py) that globs constraints/*.py?
Hooks: Do all sibling skills declare the same hooks in their YAML frontmatter? If a hook is present in some siblings but not others, is the gap justified? (Produce a Hook Coverage Matrix: skills × hooks)
Script wiring: Is every check script referenced in all three layers: (a) hook frontmatter, (b) batch orchestrator, (c) verification-checks definition? (Produce a Script Wiring Matrix: scripts × invocation points)

P08 — Constraint/convention test coverage:

Do all rules live in a single constraints/ directory? (no separate conventions/ directory)
Is the constraint/convention distinction based on presence of a .py check script, not directory location?
Does check-all.py (auto-discovering test runner) exist? Does it glob constraints/*.py — no manual wiring?
Does the verification phase run both legs: constraint checks (test runner, hard block) AND convention scoring (reviewer subagent loads .md files without .py pairs, soft block)?
Are there .md-only files (conventions) that could graduate to constraints by adding a .py check script?
Compute coverage from the filesystem: len(*.py) / len(*.md) — what percentage of rules have mechanical tests?

P09 — Iteration strategy:

Does each phase have an appropriate iteration topology? (one-shot, serial, parallel, team)
Are exit conditions structural (tests, convergence, human approval) not honor-system (promises)?

P10 — Post-subagent enforcement (from dev-debug v5.0 audit, March 16 2026):

When a subagent returns, what is main chat allowed to do? Is there an explicit tool whitelist?
Is "verification" defined concretely for this domain? (Without a definition, investigation gets disguised as verification)
Are operational tools (Bash commands beyond test running, Read on source files, Grep/Glob) restricted after subagent returns?
Is there a topic change protocol? (Without one, off-topic user messages silently kill iterative loops)

The post-subagent moment is the highest-risk point in any delegated workflow. If the audit finds no enforcement there, flag it as a critical gap.

P11 — Deviation rules (from GSD 4-rule system):

Do implementation phases have a deviation rule system (auto-fix for bugs/missing/blocking, STOP for architectural)?
Are deviation categories adapted to the domain?
Are deviations tracked and summarized per task?

P12 — State management:

Does the workflow use .planning/ for state files (not .claude/ or scattered locations)?
Are standard state files present (.planning/SPEC.md, .planning/PLAN.md, .planning/STATE.md, .planning/LEARNINGS.md)?
Is state file-based, git-trackable, and human-editable?

P13 — Session handoff:

Does the entry point check for .planning/HANDOFF.md on startup?
Is the handoff document structured with frontmatter and mandatory sections?
Can work resume from a handoff without re-discovering context?

P14 — Checkpoint types:

Are gates classified by type (human-verify, decision, human-action)?
Can the workflow auto-advance human-verify checkpoints in autonomous mode?
Are true decision points (multiple valid approaches) distinguished from rubber-stamp approvals?

P15 — Context monitoring:

Do phases check context availability before starting expensive work?
Is there a handoff trigger when context is low (≤35%)?
Does the workflow degrade gracefully or just produce garbage at context exhaustion?

P16 — Summary frontmatter:

Do phase completions produce structured YAML summaries?
Do summaries include implements, requires, provides, affects fields?
Is the one-liner substantive (not "Phase complete")?

P17 — Agent tool restrictions:

Are verification/review agents restricted to read-only tools via allowed-tools frontmatter?
Can a verifier Write or Edit? (it shouldn't — that bypasses plan-execute-verify)
Are tool restriction tiers appropriate for each agent role?

P18 — Requirement traceability:

Do requirements have unique IDs in .planning/SPEC.md (e.g., AUTH-01)?
Do .planning/PLAN.md tasks reference requirement IDs?
Does .planning/VALIDATION.md map every ID to test evidence?
Is there a scope classification (v1/v2/out-of-scope)?

P19 — Autonomous phase chaining:

Can phases chain automatically without human intervention at every step?
Does the workflow batch ambiguities (smart discuss) instead of sequential asks?
Does it re-read the plan after each phase to catch dynamically inserted phases?
Are blockers handled with retry/skip/stop options?

P19b — Visual output for human verification:

Do decision checkpoints offer visual artifacts when the human's review pattern suggests them?
Does the workflow log what the human actually looks at during review (in .planning/LEARNINGS.md)?
If the human has asked for the same view 3+ times, has it been automated into a script?

P20 — Hooks over prompt enforcement:

Are mechanically-checkable constraints enforced via scoped hooks (PreToolUse/PostToolUse in skill frontmatter)?
Or are they enforced only via prompt text (Iron Laws, Red Flags) that consume context and can be rationalized away?
Specifically check for: phase gate enforcement (prerequisite artifact checks), file extension guards, path guards, tool parameter validation, tool sequence enforcement, post-subagent restrictions
Behavioral/motivational constraints (rationalization tables, drive-aligned framing) should STAY as prompt — hooks can't teach reasoning
Score based on: how many mechanical constraints are prompt-only when they could be hooks?

P21 — Auto-loader usage for constraints:

Do phase skills that load constraint prose use the bang-invoked auto-loader?

!`uv run python3 ${CLAUDE_SKILL_DIR}/../../scripts/load-constraints.py skill-name`

Or do they list Read() calls for each constraint .md file manually?
Why this matters: The auto-loader + applies-to frontmatter is the wiring that makes atomic constraints work. Manual Read() lists mean adding a new constraint requires editing every skill that should load it — silent drift is the default failure mode.
Check: Run uv run python3 ${CLAUDE_SKILL_DIR}/../../references/constraints/auto-loader-usage.py. Every flagged SKILL.md is a violation. (Bare relative path won't resolve from the auditor's CWD — always prefix with ${CLAUDE_SKILL_DIR}/../../.)
Exceptions: router skills that immediately delegate (no constraint evaluation), ad-hoc single-file references (not phase sets), plugins without scripts/load-constraints.py.
Score: count of phase skills using the loader / count of phase skills that load ≥2 constraints. Below 80% = critical gap.

Gate: Architecture Scored [checkpoint: human-verify, auto-advanceable]

Verify scores for all P01-P21 (+ P19b) principles are present
Each principle must have numeric score (0-10) + 1-line justification
If any principle ID is missing, score it now
Composite = average of scored (non-N/A) principles

Update .planning/wc/{name}/STATE.md:

step: 2-score
status: completed
implements: [WC-09]
requires: [all workflow skill files]
provides: [P01-P21 scores with justifications]
affects: [.planning/wc/{name}/STATE.md]
one-liner: "P01-P21 scored with line-number evidence by the wc-audit reviewers; composite computed in JS."

Proceed to Step 3. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 3: Score Against Enforcement Checklist

!cat ${CLAUDE_SKILL_DIR}/../../references/enforcement-checklist.md You MUST read this file before scoring. No scoring from memory.

For each of the 13 patterns, score:

Present - pattern exists and is well-implemented
Weak - pattern exists but is insufficient (e.g., soft language instead of Iron Law)
Absent - pattern is missing where it should exist

Identify the highest-drift phases with the weakest enforcement - these are the critical gaps.

Gate: Enforcement Scored [checkpoint: human-verify, auto-advanceable]

Verify all 13 patterns were scored
Each pattern must be marked: Present / Weak / Absent
If any pattern is missing, score it now

Update .planning/wc/{name}/STATE.md:

step: 3-enforcement
status: completed
requires: [enforcement-checklist.md, all workflow skill files]
provides: [13-pattern scores per phase]
affects: [.planning/wc/{name}/STATE.md]
one-liner: "13 enforcement patterns scored Present/Weak/Absent per phase; weakest high-drift phases flagged."

Proceed to Step 3b. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 3b: Audit Path Portability

Skills run in the user's project CWD, not the plugin directory. Every path in a SKILL.md that references plugin-internal files must resolve regardless of CWD.

Scan every SKILL.md and references/*.md file in the workflow for these patterns:

Relative script paths — uv run python3 scripts/, uv run python3 ../, uv run python3 ../../ referencing plugin scripts
- These break because the agent's CWD is the user's project
- Fix: Use ${CLAUDE_SKILL_DIR}/../.. for absolute paths:
```
uv run python3 "${CLAUDE_SKILL_DIR}/../../skills/SKILL/scripts/script.py" args
```
- Or use ${CLAUDE_SKILL_DIR} for files within the same skill directory:
```
uv run python3 "${CLAUDE_SKILL_DIR}/scripts/script.py" args
```
Relative Read() paths — Read("../../skills/..."), Read("../audit-check/SKILL.md")
- The Read tool requires absolute paths; ../../ resolves from user's project CWD, not skill directory
- Fix: Use ${CLAUDE_SKILL_DIR}/../.. or ${CLAUDE_SKILL_DIR}:
```
Read `${CLAUDE_SKILL_DIR}/../../skills/SKILL-NAME/SKILL.md` and follow its instructions.
```
Dynamic context via bang-backtick injection — For constraint files that should be inlined at skill load time, use the pattern: exclamation mark followed by backtick-cat path backtick. Example: BANG + `cat ${CLAUDE_SKILL_DIR}/../../references/file.md`. This inlines the file contents at skill load time. Note: bang-backtick injection only works in top-level skills loaded via Skill(). Internal skills loaded via Read() should use direct Read() instructions instead.
Path variable substitution — the two variables apply to DIFFERENT contexts:

Hook-command variable misuse — the April 2026 incident ⚠️
- What happened: course-materials plugin v2.83.1 (Apr 15 2026) switched hook commands from ${CLAUDE_PLUGIN_ROOT}/hooks/... to ${CLAUDE_SKILL_DIR}/../../hooks/... based on a misdiagnosis claiming CLAUDE_PLUGIN_ROOT was "unset at PreToolUse runtime". Five skills across two plugins (teaching + derivative exam skills) shipped with broken hook paths for 9 days.
- Why it stayed hidden: the affected hooks (no-agent-resume-guard.py, context-monitor.py) default to {"decision": "approve"} when the script runs. When the script doesn't exist at all, Claude Code also approves. The silent-failure mode was indistinguishable from a clean approval, so "no enforcement" looked like "approved every time".
- Why exam-prep finally exposed it: a new skill added matcher: "*" under PreToolUse for context-monitor.py. matcher: "*" fires on every tool call in the session, including tool calls before the new skill was ever invoked via Skill(). At that moment ${CLAUDE_SKILL_DIR} was empty, producing a nonexistent path — Claude Code blocked every tool call with the hook's failure.
- The lesson: ${CLAUDE_SKILL_DIR} in hook frontmatter is a silent-failure landmine. It appears to work because existing hooks default-approve. Add one blocking hook — or a broad matcher — and the whole plugin surfaces the latent bug at once.

Hook Command Variable Audit (mandatory during Path Portability review):

# This command should return EMPTY — any hit is a defect:
grep -rn "command:.*\${CLAUDE_SKILL_DIR}" skills/*/SKILL.md

# All hook commands should match this pattern:
grep -rn "command:.*uv run python3 \${CLAUDE_PLUGIN_ROOT}" skills/*/SKILL.md

If the first grep returns anything, flag as a Critical Gap.

Score:

Clean — no broken paths found AND no ${CLAUDE_SKILL_DIR} in hook command fields
Partial — some paths fixed, others remain
Broken — relative paths present in skill instructions OR ${CLAUDE_SKILL_DIR} in hook command fields (even if file paths happen to resolve when tested)

Gate: Path Portability Scored [checkpoint: human-verify, auto-advanceable]

Verify all SKILL.md and references/*.md files were scanned
Every uv run python3 ../ and Read("../ pattern was flagged
Score is recorded

Update .planning/wc/{name}/STATE.md:

step: 3b-portability
status: completed
requires: [all SKILL.md and references/*.md files]
provides: [path portability score]
affects: [.planning/wc/{name}/STATE.md]
one-liner: "Path portability scored Clean/Partial/Broken; hook-command variable audit run; candidacy scan fed into Step 4."

Ultracode-Workflow Candidacy Scan (feeds Step 4 Recommendations — no separate gate)

Review fan-out — N read-only agents (per section/lecture/question/source/footnote) → computed gate / structured findings.
Write/transform fan-out — N write-agents (per file/site/lecture/section) that create or transform artifacts from a fixed spec (codemod, migration, per-item spec-driven generation), worktree-isolated. The docs' flagship case: 500-file migration, "make the change." Do not skip these — they are often the strongest candidates.

Flag a phase when the SHAPE qualifies AND it wins on ≥1 value driver:

Shape (required): the phase dispatches N agents "one per X" over a known list, and either the skill consumes their aggregated results (review) OR each does an independent per-item mutation (write). A numeric gate is NOT required.
At least one value driver: (a) parallelism, (b) context isolation, (c) a deterministic gate replacing a model-reported "recompute by hand" score (strongest review signal), or (d) independent per-item mutation at scale (write fan-out — migrations, codemods, per-item generation).

Proceed to Step 4. (STATE.md step-chain hook enforces this transition — update STATE.md before advancing.)

Step 4: Output Audit Report

The format below documents what result.reportMarkdown contains (so you can sanity-check the workflow's output) — it is the spec the workflow renders to, not a worksheet to fill in yourself.

Format:

## Audit: [Workflow Name]

### Architecture Scores (P01-P21)
| ID | Principle | Score | Notes |
|----|-----------|-------|-------|
| P01 | Phased decomposition | [0-10] | [notes] |
| P02 | Gates (deterministic/judgment) | [0-10] | [notes] |
| P03 | Structural gate enforcement | [0-10] | [notes] (STRUCTURAL/total) |
| P04 | Independent verification | [0-10] | [notes] |
| P05 | Artifact review (4-level) | [0-10] | [notes] |
| P06 | Two entry points | [0-10] | [notes] |
| P07 | Cross-skill consistency | [0-10] | [notes] |
| P08 | Constraint/convention coverage | [0-10] | [notes] |
| P09 | Iteration strategy | [0-10] | [notes] |
| P10 | Post-subagent enforcement | [0-10] | [notes] |
| P11 | Deviation rules | [0-10] | [notes] |
| P12 | State management | [0-10] | [notes] |
| P13 | Session handoff | [0-10] | [notes] |
| P14 | Checkpoint types | [0-10] | [notes] |
| P15 | Context monitoring | [0-10] | [notes] |
| P16 | Summary frontmatter | [0-10] | [notes] |
| P17 | Agent tool restrictions | [0-10] | [notes] |
| P18 | Requirement traceability | [0-10] | [notes] |
| P19 | Autonomous phase chaining | [0-10] | [notes] |
| P19b | Visual output | [0-10] | [notes] |
| P20 | Hooks over prompt | [0-10] | [notes] |
| P21 | Auto-loader usage | [0-10] | [notes] (loader skills / phase skills ≥2 constraints) |

### Gate Enforcement Matrix
| Transition | Gate | Artifact | Producer Writes? | Consumer Checks? | Hook Enforced? | Status |
|------------|------|----------|-------------------|-------------------|----------------|--------|
| [phase A] → [phase B] | [gate desc] | [artifact file] | ✅/❌ | ✅/❌ | ✅/❌ | HOOK/STRUCTURAL/ADVISORY ⚠️ |

### Enforcement Coverage
| Pattern | Phase 1 | Phase 2 | ... | Phase N |
|---------|---------|---------|-----|---------|
| Iron Laws | ✅/⚠️/❌ | ... | ... | ... |
| ... | ... | ... | ... | ... |

### Path Portability
| File | Pattern | Status |
|------|---------|--------|
| skills/X/SKILL.md | `uv run python3 scripts/foo.py` | ❌ Broken / ✅ Fixed |
| skills/Y/SKILL.md | `Read("../../lib/...")` | ❌ Broken / ✅ Fixed |

### Ultracode-Workflow Migration Candidates
| Phase | Fan-out? | Worker mode (review/transform) | Value driver | Recommend migrate? (strong/moderate) | Note |
|-------|----------|--------------------------------|--------------|--------------------------------------|------|
| [phase] | ✅/❌ (one per X) | review / transform / — | parallelism / context / gate / per-item-mutation | ✅ Mode 3 (strong\|moderate) / ❌ leave | [why] |

(From the Ultracode-Workflow Candidacy Scan. "no ultracode-workflow candidates" if none qualify.)

### Critical Gaps
1. [Highest priority gap + recommendation]
2. [Second priority gap + recommendation]
...

### Recommendations
[Specific, actionable changes]

Render score trend (if SCORES.md exists from a prior audit):

uv run python3 ${CLAUDE_SKILL_DIR}/../../scripts/render-audit-scores.py .planning/wc/{name}/SCORES.md

Gate: Audit Reported [checkpoint: human-verify, auto-advanceable]

.planning/wc/{name}/AUDIT.md exists and contains result.reportMarkdown (P01-P21 table, enforcement coverage, path portability, candidacy table, critical gaps)
result.composite was appended to .planning/wc/{name}/SCORES.md
.planning/wc/{name}/VALIDATION.md maps the WC-NN evidence (P18 self-application)
The gate verdict presented to the user is result.overallPass verbatim — not recomputed

step: 4-report
status: completed
implements: [WC-09]
requires: [all workflow skill files]
provides: [AUDIT.md, VALIDATION.md]
affects: [.planning/wc/{name}/AUDIT.md, .planning/wc/{name}/VALIDATION.md]
one-liner: "AUDIT.md rendered from the wc-audit result (composite + matrices + candidacy); VALIDATION.md maps WC-NN evidence."

<EXTREMELY-IMPORTANT> ### The Iron Law of Thorough Scoring

NO PRINCIPLE SCORE WITHOUT LINE-NUMBER EVIDENCE. This is not negotiable.

If you cannot point to a specific line, file, or pattern that justifies the score — the score is wrong. </EXTREMELY-IMPORTANT>

Deviation Rules for Mode 2 (Auditing)

During auditing, unplanned issues may arise. Apply these deviation rules:

Priority: R4 (STOP) > R1-R3 (auto) > unsure → R4

Mode 2 Enforcement

Rationalization Table — Mode 2

Red Flags — STOP If You Catch Yourself:

Staged Review — Mode 2

Delete & Restart — Mode 2

Drive-Aligned Framing — Mode 2

Mode 3: Improve Workflow

<EXTREMELY-IMPORTANT> ## The Iron Law of Workflow Improvement

MODE 3 IS AN AUDIT-FIX LOOP. THE SUBSTRATE GATE DECIDES WHEN TO STOP, NOT YOU AND NOT A BARE COMPOSITE NUMBER. This is not negotiable.

The structural problem Mode 3 solves (two failure modes, opposite directions):

Stopping too early by self-declaring "diminishing returns" before the substrate is clean. The loop + /goal evaluator prevents this — you cannot stop while a critical or an Absent enforcement pattern remains.
Chasing a noisy proxy forever. The composite is an LLM panel: it re-rolls ±0.2 each run, its domain-ceiling denominator drifts, and it regenerates new minor findings every pass. Targeting composite ≥ 9.5 sends the loop onto a treadmill where fixing 4 findings surfaces 4 new ones and the number never converges (empirically every workflow asymptotes ~9.0; see project_wc_mode3_asymptote). The last 0.5 to 9.5 can only be bought by over-enforcing creative/low-drift phases — which violates the Step-4 drift tiering and the no-speculative-enforcement rule. So 9.5 rewards making the workflow worse. </EXTREMELY-IMPORTANT>

Why the gate is the substrate, not 9.5 (calibrated ceiling)

Step 1: Run Initial Audit (Mode 2)

State initialization: Create .planning/wc/{name}/STATE.md with this YAML template:

---
mode: improve
step: 1-initial-audit
status: in_progress
target: [workflow name]
implements: [WC-12]
requires: [target workflow files]
provides: [baseline composite]
affects: [.planning/wc/{name}/STATE.md, .planning/wc/{name}/SCORES.md]
one-liner: "Improve loop started on {target} — running baseline wc-audit before the /goal climb."
---

Context monitoring: Mode 3 runs multi-iteration audit-fix loops. Each iteration consumes significant context. Check availability:

If context is low (≤35% remaining), write .planning/wc/{name}/HANDOFF.md with current iteration, score, and remaining gaps before starting the next iteration.
If context is critical (≤25% remaining), write HANDOFF.md immediately — do not start another iteration.

Run Mode 2 on the target workflow. This produces the baseline score.

Gate: Mode 2 audit report exists with numeric scores for all P01-P21 principles. [checkpoint: human-verify, auto-advanceable]

Step 2: Launch Audit-Fix Loop

/goal Workflow [WORKFLOW_NAME] reaches result.substratePass=true (0 critical, no enforcement Absent, portability Clean) AND composite >= 9.0 AND composite flat (within ±0.2 of the prior SCORES.md row). Stop after 10 turns. Do NOT keep iterating to lift the composite once the substrate gate is clean and the composite is flat — that is the treadmill.

Before launching the goal, persist the second link of the improve chain (the hook's improve chain is 1-initial-audit → 1-audit-loop):

step: 1-audit-loop
status: in_progress
requires: [Mode 2 audit report]
provides: [score-gated fix iterations]
affects: [.planning/wc/{name}/STATE.md, .planning/wc/{name}/SCORES.md, target workflow files]
one-liner: "/goal audit-fix loop launched (gate: substratePass + composite≥9.0 + flat); each turn re-runs wc-audit and appends composite + substratePass to SCORES.md."

Each turn under the active goal follows this exact sequence:

Phase A: AUDIT ──→ Phase B: DECIDE ──→ Phase C: FIX
  [fresh subagent]   [check score]       [targeted edits]
    │                    │                    │
    ▼                    ▼                    ▼
  AUDIT.md         substratePass &&      Fix gaps by priority:
  SCORES.md        composite≥9.0 &&       1. criticals (any severity:critical)
                   flat(±0.2)?            2. enforcement Absent → Present
                        │                 3. portability → Clean
                    YES ──→ end turn;      4. then P<9 medium gaps
                           /goal evaluator      │
                           marks done           ▼
                        │                   End turn → /goal
                    NO ──→ Phase C          refires Phase A

Phase A: AUDIT (the wc-audit workflow — MANDATORY, independent by construction)

WF=$(command ls -d ~/.claude/plugins/cache/edwinhu-plugins/workflows/*/workflows/wc-audit.js 2>/dev/null | sort -V | tail -1)
[ -z "$WF" ] && WF="${CLAUDE_SKILL_DIR}/../../workflows/wc-audit.js"

# Full audit on iteration 1; selective re-audit thereafter:
Workflow({ scriptPath: "<WF>", args: {
  targetWorkflow: "{name}", projectDir: "<abs repo root>", pluginRoot: "<abs .../workflows dir>",
  threshold: 9.0,                                   // calibrated ceiling, NOT 9.5 — composite is advisory; substratePass is the gate
  onlyChecks: <prev result.reviewersThatFlagged>,   // omit on iteration 1
  priorReviews: <prev result.reviews>               // omit on iteration 1
} })

<EXTREMELY-IMPORTANT> **THE AUDIT IS THE wc-audit WORKFLOW, NOT YOUR OWN RE-READ. If you score your own fixes by hand, you are rubber-stamping.**

Phase B: DECIDE `[checkpoint: decision]`

Read .planning/wc/{name}/SCORES.md. Render the score trend for visual context:

uv run python3 ${CLAUDE_SKILL_DIR}/../../scripts/render-audit-scores.py .planning/wc/{name}/SCORES.md

Check the substrate gate, then the calibrated composite + flatness. "Flat" = this iteration's composite is within ±0.2 of the prior SCORES.md row.

Phase C: FIX

Post-subagent boundary (after wc-audit returns). The audit is the wc-audit workflow's job; fixing is yours. Keep the two separate:

After fixing, you re-run the workflow — you do NOT substitute your own read for the auditor's verdict.

Address findings from .planning/wc/{name}/AUDIT.md, prioritized by severity:

Fix all principles scoring < 7.0 first (critical gaps)
Then principles scoring 7.0-8.9 (medium gaps)
Then principles scoring 9.0-9.4 (polish)

Fix reference — common gap → fix mapping:

Deviation Rules for Mode 3 Phase C (Fixing)

During fix application, unplanned issues may arise. Apply these deviation rules:

Priority: R4 (STOP) > R1-R3 (auto) > unsure → R4

Fix rules:

Targeted changes only — do NOT rewrite entire skill files
Each fix addresses ONE gap from the audit
After fixing, do NOT self-assess — the next iteration's audit will judge
End your turn immediately so the loop feeds you back for re-audit. Do NOT ask "should I continue?", do NOT summarize the fixes you just made, do NOT wait for confirmation — the /goal evaluator re-fires Phase A on its own. Pausing between fix iterations is procrastination disguised as courtesy; it strands the loop.

Efficiency Optimizations

Mode 3 can be expensive. These optimizations reduce cost without sacrificing audit independence.

1. Batch Fixes Per File

Group all fixes targeting the same file into a single edit. Don't make 5 separate edits to the same constraint file — read the audit findings, plan all changes, apply them in one Edit call.

2. Scoped Re-Audit After Iteration 1

3. Prioritize Cheapest High-Impact Fixes

After the audit, sort gaps by impact / effort:

Adding a .md + .py pair to constraints/ (all skills inherit, auto-discovered) = high impact, low effort
Adding allowed-tools frontmatter to 3 reviewer skills = high impact, 5 seconds each
Rewriting a phase skill's entire gate structure = medium impact, high effort

Fix the cheap high-impact gaps first. This maximizes score improvement per iteration.

4. Domain-Appropriate Scoring

Some principles have natural ceilings in certain domains:

Writing gates are judgment-based (not deterministic) — 9.0 is the natural ceiling for "gates" in writing
Writing has one midpoint because the domain only needs one — 9.0 is appropriate for "two entry points"

Caution: This is the auditor's call, not the fixer's. The fixer cannot declare a domain ceiling to avoid work. Only the independent auditor can classify a principle as domain-limited.

Why This Must Be a Loop (Not Manual Iteration)

Rationalization Table — Mode 3

Red Flags — STOP If You Catch Yourself:

Drive-Aligned Framing — Mode 3

| Your Drive | Why You Cut Corners | What Actually Happens | The Drive You Failed | |------------|--------------------|-----------------------|---------------------| | Helpfulness | "Ship the fixes faster, skip the re-audit" | Unverified fixes ship broken enforcement; the user discovers a substrate gap in production. | Anti-helpful | | Competence | "I can tell from the diff this fixes the gap" | Self-assessment is rubber-stamping; the auditor exists because you can't see your blind spots. | Incompetent | | Efficiency | "Re-reading all files each iteration is wasteful" | Fresh-context audit catches regressions incremental review misses. | Anti-efficient | | Approval | "I'll grind to 9.5 because a bigger number looks better to the user" | You burn tokens on a treadmill and over-enforce creative steps, making the workflow worse. A flat 8.7 with substratePass beats a gamed 9.5. | Lost approval | | Honesty | "I'll call it 9.5-equivalent" or "I'll ignore the open critical" | Misrepresenting a noisy proxy, or shipping with a real substrate blocker, is fabricating completion. | Dishonest |

<EXTREMELY-IMPORTANT> ## Iron Laws of Workflow Creation

NO WORKFLOW WITHOUT PHILOSOPHY

Every workflow must trace back to PHILOSOPHY.md. If you can't explain how a phase serves phased decomposition, gates, or adversarial review, the phase doesn't belong.

NO PHASE WITHOUT A GATE

Every phase needs a gate — deterministic (test passes, file exists) or judgment-based (agent/human evaluates quality). Use the strongest gate available for the domain. No gate = not a real phase.

NO HIGH-DRIFT PHASE WITHOUT ENFORCEMENT

Identify where the agent is most tempted to shortcut. Enforce hardest there. Implementation and verification phases ALWAYS need Iron Laws.

NO UNREVIEWED ARTIFACT CROSSING A PHASE BOUNDARY

NO SKILL FAMILY WITHOUT SHARED ENFORCEMENT

NO CONSTRAINT WITHOUT A CO-LOCATED CHECK SCRIPT

NO VERIFICATION WITHOUT BOTH LEGS

NO VERIFIER WITH WRITE ACCESS

NO LONG WORKFLOW WITHOUT CONTEXT MONITORING

NO NESTED AGENT DISPATCH (Iron Law of Flat Dispatch)

BAD:  orchestrator → dispatcher agent → 5× sub-sub-agents (results lost)
GOOD: orchestrator → 5× agents directly in parallel (all return reliably)

Red Flags - STOP If You Catch Yourself:

Rationalization Table

Why Skipping Steps Hurts the Thing You Care About Most

You skip steps because you think it's helpful, efficient, or competent. Here's what actually happens:

The protocol is not overhead you pay. It is the service you provide.

Every time you skip steps to "deliver faster," you choose YOUR comfort over the USER's outcome. The user doesn't experience your tedium—they experience your workflow's failure rate.

Adoption

edwinhu/workflow-creator

$ install --global

Security Scan Results

SKILL.md

Startup: State Check

Mode 1: Create New Workflow

Step 1: Ground in Philosophy

Step 2: Interview

INTERVIEW.md Review Gate

Step 3: Propose Phase Decomposition

Structural Gate Artifacts

Hook-Enforced Gates (Preferred Pattern)

Iteration Topology

Verification Depth

Test Gap Validation Phase

Checkpoint Types

Context Monitoring

Summary Frontmatter

Agent Tool Restrictions (READ-ONLY Verifiers)

Requirement Traceability

Autonomous Phase Chaining

Step 3b: Add Artifact Review Gates

DESIGN.md Review Gate

Step 4: Apply Enforcement Patterns

Hooks Over Prompt Enforcement

Deviation Rules for Implementation Phases

Step 4b: Common Enforcement Across Skill Families

Layer 1: Shared Constraints (Co-located in constraints/)

Constraint Architecture: Co-located Pairs, Auto-discovered

The One Directory

Rule File Structure

Check Script Interface

Test Runner: Auto-discovery

Verification Architecture: Two Legs

Layer 2: Hook Coverage (Structural Enforcement)

Layer 3: Script Wiring (Auto-discovery + Hooks)

Step 5: Design Two Entry Points

Midpoint Constraint Loading

Shared Constraint Files

Session Handoff Support

Step 6: Generate Workflow Files

State Folder Convention

Visual Output for Human Verification (Learn-by-Doing)

Optional: generate the files with the wc-generate transform workflow

Deviation Rules for Steps 4-6 (File Generation)

Delete & Restart Protocol

Step 7: Self-Audit the Generated Workflow

Post-Subagent Enforcement (Step 7)

Mode 2: Audit Existing Workflow

Step 1: Read the Workflow

Step 2: Score Against Core Principles (P01-P21)

Step 3: Score Against Enforcement Checklist

Step 3b: Audit Path Portability

Ultracode-Workflow Candidacy Scan (feeds Step 4 Recommendations — no separate gate)

Step 4: Output Audit Report

Deviation Rules for Mode 2 (Auditing)

Mode 2 Enforcement

Rationalization Table — Mode 2

Red Flags — STOP If You Catch Yourself:

Staged Review — Mode 2

Delete & Restart — Mode 2

Drive-Aligned Framing — Mode 2

Mode 3: Improve Workflow

Why the gate is the substrate, not 9.5 (calibrated ceiling)

Step 1: Run Initial Audit (Mode 2)

Step 2: Launch Audit-Fix Loop

Phase A: AUDIT (the wc-audit workflow — MANDATORY, independent by construction)

Phase B: DECIDE [checkpoint: decision]

Phase C: FIX

Deviation Rules for Mode 3 Phase C (Fixing)

Efficiency Optimizations

1. Batch Fixes Per File

2. Scoped Re-Audit After Iteration 1

3. Prioritize Cheapest High-Impact Fixes

4. Domain-Appropriate Scoring

Why This Must Be a Loop (Not Manual Iteration)

Rationalization Table — Mode 3

Red Flags — STOP If You Catch Yourself:

Drive-Aligned Framing — Mode 3

Layer 1: Shared Constraints (Co-located in `constraints/`)

Phase B: DECIDE `[checkpoint: decision]`

Layer 1: Shared Constraints (Co-located in `constraints/`)