lidge-jun/goal

Goal Execution Skill

Guidelines for AI-driven goal work within cli-jaw. Applies when a goal is active via /goal or cli-jaw goal set, and especially when using /goal plan for AI-driven goal formulation.

Two Modes

Mode 1: Direct Goal (/goal "objective")

User provides the objective directly. AI executes following these guidelines. The objective is taken verbatim as the goal.

Mode 2: AI-Planned Goal (/goal plan "objective")

AI analyzes the user's intent and formulates a specific, long-term executable goal. The user's input is a direction, not the final objective.

Planning Process (fully autonomous — no user confirmation step):

1. Context Scan: Read project docs (README, CLAUDE.md, structure/, devlog/), recent git history, and relevant source code
2. Scope Analysis: What does the objective mean in this specific codebase? What subsystems are affected?
3. Dependency Map: What existing modules, APIs, data flows, or external services are involved?
4. Risk Assessment: What could go wrong? Integration risks? Breaking changes? Performance impact?
5. Sub-goal Decomposition: Break into PABCD-compatible work units, each independently verifiable
6. Execution Plan: Formulate a specific, measurable goal with clear completion criteria
7. Refine & Execute: Immediately refine the goal objective via API and enter PABCD autonomously

Core Guidelines

1. PABCD Integration

Every goal should be executable through the P-A-B-C-D orchestration cycle:

• P (Plan): Write a complete plan with diff-level precision
• A (Audit): Dispatch an employee to audit the plan against real code
• B (Build): Implement the plan; dispatch verification employees
• C (Check): Final sanity check (tsc, consistency, structure docs)
• D (Done): Summarize and return to IDLE

When a goal is active, it supersedes PABCD phase gates for autonomous execution. Phase transitions are mandatory shell commands (cli-jaw orchestrate A/B/C/D), not status text. "Supersedes gates" means no user wait — it does NOT mean skip P/A/B/C/D or merge work-phases into B. The evidence attestation (--attest) is still required in goal mode: it is proof-of-work, not a user-approval wait, so each forward transition must carry a real {"from","to","did"} (C→D also checkOutput/exitCode).

Multi-phase / loop goals: each sub-goal / work-phase = one FULL PABCD cycle (P→A→B→C→D). After D (state → IDLE), run cli-jaw orchestrate P to start the next work-phase; repeat until the objective is met. Never run B for several work-phases back-to-back. A "loop"/"루프" goal is multi-pass: pre-plan the full slice map and WRITE all per-phase decade docs to diff-level up front (DIFFLEVEL-ROADMAP-01) — scaffolding empty stubs is not pre-planning. It may open with a design-only PABCD pass (Phase 0) before the first implementation work-phase. Faithfully execute each PABCD-phase — never rubber-stamp a phase to advance.

2. Documentation Workflow

Documentation is required throughout the goal lifecycle:

• Planning: Create devlog entries at devlog/_plan/<YYMMDD_topic>/ using decade numbering:
  • 00-09: Research, specs, analysis (mandatory)
  • 10-19: Phase 1 artifacts
  • 20-29: Phase 2 artifacts
  • 30+: Additional phases
• Progress: Use cli-jaw goal update "<summary>" --evidence "<path>" at every milestone
• Bug tracking: Document bugs and fixes explicitly in devlog entries
• Phase evidence: Every PABCD phase must produce documentation evidence (devlog path or structure update)

3. Commit Discipline

• Small, atomic commits after each logical change
• Never batch multiple unrelated changes into one commit
• Each commit must be self-contained and independently reversible
• Commit messages should explain WHY, not WHAT
• Do not commit generated files, secrets, or temporary state

4. Verification Process

Scale verification effort to change complexity (adapted from harness verification-tiers pattern):

| Tier | Criteria | Who Verifies | Evidence Required | |------|----------|--------------|-------------------| | LIGHT | <5 files, <100 lines changed | CLI sub-agent | LSP diagnostics clean | | STANDARD | Default for mid-size changes | jaw employee dispatch | Diagnostics + build pass | | THOROUGH | >20 files OR security/architectural changes | jaw employee (full review) | All tests pass + independent code review |

Tier Selection Logic:

• Security-related changes (auth, credentials, tokens, permissions) -> always THOROUGH
• Architectural changes (config, schema, types, package.json) -> always THOROUGH
• Explicit user request ("careful", "thorough", "critical") -> THOROUGH
• Small isolated fixes with full test coverage -> LIGHT
• Everything else -> STANDARD

Override Triggers:

• User says "quick", "simple", "trivial" -> LIGHT
• User says "thorough", "careful", "important" -> THOROUGH

5. Sub-agent/Employee Verification Mandates

• Before claiming any phase completion, dispatch independent verification
• Verification is NOT an approval gate: dispatch -> receive result -> act immediately
• Independent reviewers must challenge completion claims, not rubber-stamp
• Evidence must be authoritative: file paths, command output, test results, runtime behavior
• If verification finds issues, fix them and re-verify; do not skip

6. Quality Gates

Adapted from the ultragoal pattern:

• No placeholder evidence: Reject any evidence containing "todo", "tbd", "stub", "not implemented", or "fake pass"
• Artifact paths required: Every validation summary must include concrete file paths
• Evidence triple required for completion: Documentation evidence + Implementation evidence + Verification evidence
• Independent review for final completion: Dispatch a separate reviewer to challenge whether all requirements are truly met
• Treat completion as unproven: Default to "not done" until evidence proves otherwise

7. When Blocked

• Try alternative approaches before declaring blocked
• If genuinely blocked (need auth, hardware, human business decision):
  1. Record all progress with cli-jaw goal update
  2. Dispatch independent reviewer to confirm no viable path remains
  3. Pause with audit: cli-jaw goal pause --agent --audit "<review summary>"
• Never declare blocked merely because work is hard, slow, or uncertain
• Never auto-complete a goal; completion is reserved for explicit user request or verified evidence

Steer Prompt Templates

Direct Goal Steer

[System] User set a new goal: "<objective>" (ID: <id>). Acknowledge the goal and help the user achieve it.

Plan Goal Steer

[System] User requested AI-driven goal planning: "<objective>" (ID: <id>).
Do NOT execute the objective directly. First ANALYZE, then EXECUTE AUTONOMOUSLY:
1. Context Scan -> 2. Scope Analysis -> 3. Dependency Map -> 4. Risk Assessment -> 5. Sub-goal Decomposition -> 6. Execution Plan
After analysis, immediately refine the goal objective and enter PABCD orchestration.
Do NOT wait for user confirmation. Execute the full goal autonomously.