edwinhu/skill-creator

Skill Creator (with Superpowers Enforcement)

This skill wraps the built-in skill-creator:skill-creator with enforcement pattern awareness from the superpowers framework. It adds an enforcement audit layer to the skill-creator's draft-test-iterate loop.

When This Skill Applies

All skill creation and improvement work. This skill loads instead of the built-in skill-creator because it adds enforcement awareness that the built-in version lacks.

Process

Step 1: Classify the Skill

Before drafting, classify the skill being created:

| Type | Description | Enforcement Needs | |------|-------------|-------------------| | Workflow skill | Multi-phase process (like /dev, /ds, /writing) | High — needs Iron Laws, gates, rationalization tables | | Tool skill | Wraps a tool or API (like readwise, wrds, bluebook) | Medium — needs Red Flags for common misuse | | Knowledge skill | Domain knowledge reference (like ai-anti-patterns) | Low — needs trigger-only descriptions |

This classification determines how much enforcement audit to apply after each draft.

Anti-Patterns: Read Before Drafting

!cat ${CLAUDE_SKILL_DIR}/../../references/creator-anti-patterns.md

Step 1b: Check for Mechanical Enforcement Opportunities

Before drafting, identify what should be mechanically enforced rather than prompt-enforced. Four mechanisms are available, each resolving at a different time:

| Mechanism | Resolves at | Gives you | Use for | |-----------|------------|-----------|---------| | ${CLAUDE_SKILL_DIR} | Skill load | A path string | Script paths in Bash templates | | !command`` (bang) | Skill load | Command stdout as inline text | Injecting reference files, environment state | | Scoped hooks (Pre/PostToolUse) | Each tool call | Pass/fail gate | Mechanically checkable constraints | | SessionStart hook (once: true) | Session start | Value written to a file | Expensive computations (API calls, index builds) |

${CLAUDE_SKILL_DIR} — Script Path References

Use directly in Bash command templates — substituted at skill load time to the full absolute path:

# ✅ CORRECT: Variable substituted at load time, Claude sees literal path
uv run python3 "${CLAUDE_SKILL_DIR}/scripts/my_script.py" --arg value

# ❌ WRONG: Broken $() subshell — executes script with no args, captures garbage
SCRIPT=$(${CLAUDE_SKILL_DIR}/scripts/my_script.py) && uv run python3 "$SCRIPT" --arg value

The $() indirection pattern is a common mistake. It tries to execute the script in a subshell and capture its stdout — but scripts require arguments and fail with no args, leaving the variable empty.

Bang-Backtick Injection (!command``)

Bangs run a shell command at skill load time and inline the stdout into the prompt text. Use them to inject content, not paths:

| Use Case | Example | |----------|---------| | Inline reference files | !cat ${CLAUDE_SKILL_DIR}/../../references/constraints.md | | Environment detection | `!`if [ -f /.dockerenv ]; then echo "CONTAINER"; else echo "HOST"; fi | | Inject current state | !git branch --show-current`` |

Bangs only work in top-level skills loaded via Skill(). Internal skills loaded via Read() should use direct Read() instructions.

Scoped Hooks (PreToolUse / PostToolUse)

Hooks in skill frontmatter fire only while the skill is active — automatically cleaned up when the skill finishes:

hooks:
  PreToolUse:
    - matcher: "Write"
      hooks:
        - type: command
          command: "uv run python3 ${CLAUDE_PLUGIN_ROOT}/hooks/guard.py"
  PostToolUse:
    - matcher: "Edit|Write"
      hooks:
        - type: command
          command: "uv run python3 ${CLAUDE_PLUGIN_ROOT}/hooks/lint.py"

⚠️ Hook-command variable rule: use ${CLAUDE_PLUGIN_ROOT} in hook command: fields — not ${CLAUDE_SKILL_DIR}. The latter is for skill content (markdown body, bang-backtick commands). Hook frontmatter is a different substitution context; mixing these up causes silent failures when the hook fires outside an active Skill() session. See workflow-creator Step 3b for the April 2026 incident.

| Enforce with Hook | Keep as Prompt | |-------------------|----------------| | File path/extension guards | Fact rows (incident-grounded) | | Missing prerequisite file checks | Iron Laws with drive-consequence framing | | Tool parameter validation | Red flags (judgment-based, action-targeted) | | Post-edit lint/format checks | "Why" explanations | | Outline-before-prose guards | Deviation rule classification |

SessionStart Hooks for Expensive Resolutions

For values that genuinely require expensive computation (API calls, multi-step searches, environment probes) and are stable for the session:

hooks:
  SessionStart:
    - hooks:
        - type: command
          command: "expensive-api-call --query env > .planning/CACHED_VALUE"
          once: true

Then instruct the skill to read from .planning/CACHED_VALUE. Do NOT use for path resolution (${CLAUDE_SKILL_DIR}) or content injection (bangs) — those are free at load time. See references/sessionstart-caching.md.

The principle: if a constraint is mechanically checkable, enforce it with a hook. If it requires judgment or motivation, keep it as prompt text. Hooks cost zero tokens and can't be rationalized away.

Step 2: Invoke the Built-in Skill Creator

Use the Skill tool to invoke the built-in skill-creator:

Skill(skill="skill-creator:skill-creator")

Follow its full process: capture intent, interview, draft SKILL.md, write test cases, run evals, iterate. The built-in skill-creator handles the eval loop — do not reimplement it.

Step 3: Enforcement Audit (After Each Draft)

After writing or revising the skill draft (and before running test cases), audit it against the superpowers enforcement patterns. Read the enforcement checklist:

!cat ${CLAUDE_SKILL_DIR}/../../references/enforcement-checklist.md

Then score the draft using the process below.

For Workflow Skills (High Enforcement)

Score against all 12 patterns. Use the scoring template from the checklist. Focus on:

1. Iron Laws — Does the skill have absolute constraints for high-drift actions? Are they wrapped in <EXTREMELY-IMPORTANT> tags with strong framing? If they use soft language ("try to", "should", "consider"), they will be ignored — rewrite with action-masking language.

2. Fact Rows (supersedes Rationalization Tables, v5.36.0) — Does the skill state its incident-learned knowledge as declarative facts? Each row must be non-derivable (a number, threshold, named incident, or tool quirk from observed failures — not a restatement of the rule), with the consequence framed as a property of the action (counterproductive / unhelpful / dishonest / incompetent). Legacy excuse/reality tables in existing skills count as present but should convert on next touch; never author new ones.

3. Red Flags + STOP — Are there pattern interrupts for observable wrong actions? Must target actions ("About to X"), not intentions ("Thinking about X").

4. Gate Functions — Does every phase transition have a verifiable exit condition? "Quality is sufficient" is not a gate. "File X contains string Y" is a gate.

5. Trigger-Only Descriptions — Does the description contain ONLY trigger phrases? If it contains a process summary, the agent will follow the short description instead of reading the body. This is the single most common skill design mistake.

6. Drive-Aligned Framing — Do Iron Laws and fact rows carry helpfulness-first consequences? "Skipping X is NOT HELPFUL — [concrete user harm]" is stronger than "incorrect" or "premature" because it targets the model's strongest drive. Embed in the law or fact row itself — standalone "Your Drive | Why You Skip" tables are deprecated (they restate one consequence five times).

7. Skill Dependencies — Does each phase explicitly read and invoke the next phase? Without explicit chaining, the agent will stop and wait.

8. No Pause Between Tasks — Does the skill prevent "should I continue?" between tasks?

8b. Flat Agent Dispatch — If the skill spawns agents that perform multiple checks or tasks, does the skill spawn them ALL directly in parallel? Or does it spawn a "dispatcher" agent that spawns its own sub-agents? Three-layer delegation (skill → agent → sub-agents) fails because sub-sub-agent results don't reliably return. The orchestrator must spawn all agents directly. See workflow-creator's Iron Law of Flat Dispatch.

9. Delete & Restart — For protocol violations, does the skill mandate deletion of contaminated work?

10. Staged Review Loops — Do implementation sections have review loops with iteration limits?

11. Flowcharts as Spec — For complex processes, is there an ASCII diagram that serves as the authoritative definition?

Critical gaps = High-drift action + Absent/Weak enforcement. Fix these before running evals.

For Tool Skills (Medium Enforcement)

Score against patterns 2, 3, 5, and 10:

• Fact Rows — What are the tool's non-derivable gotchas? (e.g., "the API validates format, not correctness — an empty response returns 200"; rate limits; auth quirks)
• Red Flags + STOP — What wrong actions can the agent take? (e.g., calling a destructive API without confirmation)
• Trigger-Only Descriptions — Keep description to triggers only
• Staged Review Loops — For multi-step tool interactions, add review after each step

For Knowledge Skills (Low Enforcement)

Score against pattern 5 only:

• Trigger-Only Descriptions — This is the most important pattern for knowledge skills. If the description summarizes the knowledge, the agent reads the summary instead of the full body.

Step 4: Reconcile Tensions

The built-in skill-creator's writing advice and superpowers enforcement patterns have a genuine tension:

| skill-creator says | superpowers says | Resolution | |---|---|---| | "Explain the why, avoid heavy-handed MUSTs" | "Iron Laws use strongest framing available" | Both are right for different contexts. Use "explain the why" for standalone instructions. Use Iron Laws for high-drift actions where the agent will rationalize shortcuts. | | "Keep the prompt lean" | "Add Fact Rows, Red Flags" | Enforcement patterns go in the skill body, not the description. Progressive disclosure keeps it lean — move detailed facts to references/ if SKILL.md exceeds 500 lines. | | "Generalize from feedback, don't overfit" | "Observe failure modes, add fact rows" | Fact Rows ARE generalization. Each row captures a class of failures (the fact + its consequence), not a specific test case. |

When the built-in skill-creator suggests removing enforcement patterns because they're "not pulling their weight" or are "oppressively constrictive MUSTs," push back if the pattern addresses a real observed failure mode. The test: did an agent actually take the shortcut this pattern prevents? If yes, keep it.

Step 5: Continue the Eval Loop

Return to the built-in skill-creator's process for running test cases, grading, and iterating. After each iteration's skill revision, re-run the enforcement audit (Step 3) on the updated draft.

During the eval loop, watch for enforcement iteration signals (see "Enforcement Iteration Signals" in the anti-patterns reference loaded above).

References

• Enforcement checklist: references/enforcement-checklist.md (in plugin root) — Full 12-pattern reference with templates. Discover via: ${CLAUDE_SKILL_DIR}/../../references/enforcement-checklist.md
• Description patterns: references/skill-description-patterns.md (in plugin root) — The three description shapes (standalone / user-triggered, workflow-phase / orchestrator-triggered, internal-only) with templates and a migration guide for reclassifying an existing skill. Discover via: ${CLAUDE_SKILL_DIR}/../../references/skill-description-patterns.md
• Philosophy: PHILOSOPHY.md (in plugin root) — Three pillars (phased decomposition, deterministic gates, adversarial review). Discover via: ${CLAUDE_SKILL_DIR}/../../PHILOSOPHY.md
• Built-in skill-creator: Handles the eval loop (draft → test → grade → iterate → description optimization)