skill-audit

name

skill-audit

description

Interactive behavioral quality audit for individual skills. Evaluates against 12 quality categories (including T25 convergence loop and completion verification) to surface attention management issues and produce actionable decisions. Includes self-audit verification that all decisions were actually implemented. Uses convergence loops in its own discovery phase. Produces a decision record and updated skill files.

Document Version: 3.0 Last Updated: 2026-03-06 Status: ACTIVE

Skill Audit

Interactive, category-by-category behavioral quality audit for individual skills. Surfaces gaps in intent fidelity, workflow sequencing, prompt engineering, integration, guard rails, and user experience — then resolves each gap through guided decisions with the user.

Critical Rules (MUST follow)

1. Present ALL issues — every con, gap, and issue identified in a category MUST be presented to the user, not just the ones with multiple options.
2. Recommend with rationale — for every suggestion, state your recommendation and explain WHY. Do not present options without a recommendation.
3. Save decisions incrementally — persist to state file after each category and each phase boundary. Non-negotiable for long-running audits.
4. Crosscheck skill-creator — before implementing, verify the creator skill itself covers the gaps found in this audit.
5. Address the skill, not just the word — when behavior doesn't match intent, fix the behavioral instructions, not the naming.
6. Self-audit before completion — Phase 5 is MUST. Never skip verification. Re-read all modified files and verify every decision was implemented.
7. Mode-aware execution — Phase 1.0 selects mode (single / batch / multi) before any other work. single preserves Phase 2a per-category gating MUST. batch and multi use Phase 2b (batched findings) + Phase 2.B (decision collection). Never apply Phase 2a gating to batch/multi; never collapse single-mode per-category gates.

When to Use

• Evaluating a skill's behavioral quality before a major rewrite
• User invokes /skill-audit
• After creating a skill with /skill-creator to validate quality
• When a skill consistently produces weak output (diagnose root cause)
• Periodic quality review of critical skills

When NOT to Use

• Auditing the skill ecosystem as a whole — use /skill-ecosystem-audit
• Creating a new skill from scratch — use /skill-creator
• Quick structural validation — use npm run skills:validate

Routing Guide

• "Evaluate this skill's quality" → /skill-audit
• "Create a new skill" → /skill-creator
• "Update an existing skill" → /skill-creator (or /skill-audit first)
• "Check all skills for issues" → /skill-ecosystem-audit
• "Quick structural check" → npm run skills:validate

---

Process Overview

Phase 1: Preparation    → Validate target, read skill + standards, initialize state
Phase 2: Category Audit → 12 categories (incl. T25 + Completion Verification), interactive, decisions saved per category
Phase 2.5: Operational Deps → Scripts, hooks, data files, npm scripts, state files
Phase 3: Crosscheck     → Verify skill-creator + ecosystem impact + adjacent contracts
Phase 4: Implementation → Apply decisions (priority order, batch related changes)
Phase 5: Self-Audit     → Verify decisions implemented, process followed
Phase 6: Learning Loop  → Process feedback, invocation tracking, closure

---

Phase 1.0: Mode Selection (MUST — first gate)

Before any other work, prompt the user for mode:

Mode?
  (1) single — interactive, one category at a time (default behavior, MUST gated)
  (2) batch  — single skill, all 12 categories produced at once to tmp file
  (3) multi  — multiple skills, batched per skill, Shape Y orchestration

Record the chosen mode to state file mode field. If mode=multi, follow-up prompt: "Skills to audit (comma-separated list):" — record as skills_in_batch on the parent batch state file (see .claude/skills/skill-audit/REFERENCE.md §Parent Batch State Schema).

Phase routing after Phase 1:

• mode=single → standard flow through Phase 2a (gated, interactive)
• mode=batch → Phase 2b (batched findings) → Phase 2.B (decision collection) → Phase 3
• mode=multi → Phase 2b per skill → Phase 2.A (cross-skill patterns) → Phase 2.B (decision collection per skill) → batched Phase 3

Done when: Mode recorded to state file; for multi, skills list captured.

---

Phase 1: Preparation (MUST)

1. Validate target (MUST) — if SKILL.md doesn't exist, report error and suggest /skill-creator. Note companion file count in warm-up.
2. Read target skill's SKILL.md and companion files (MUST)
3. Read SKILL_STANDARDS.md for quality checklist (MUST)
4. Check previous audit (SHOULD) — look for existing state file. If found, include previous scores and top findings in warm-up.
5. Audit health self-check (MUST) — verify REFERENCE.md has all 10 categories, state file path is writable, SKILL_STANDARDS.md is readable.
6. Initialize state file (MUST) — path: .claude/state/task-skill-audit-{skill-name}.state.json
7. Present warm-up summary (MUST):

Skill Audit: [skill-name]
Lines: [N] | Companions: [N] | Neighbors: [list] | Last Updated: [date]
Categories: 10 | Estimated decisions: [N]
[If repeat: Previous: [date], score [N/100]. Top findings: [list]]

Done when: Warm-up presented, state initialized, user confirms proceed.

---

Phase 2a: Category Audit — single mode (Interactive, Gated, MUST)

Mode scope: This phase applies ONLY when mode=single. For mode=batch or mode=multi, skip to Phase 2b below.

Read .claude/skills/skill-audit/REFERENCE.md for the 12 category definitions, question banks, scoring rubrics, and presentation format. Category 11 (Convergence Loop Integration) evaluates T25 compliance. Category 12 (Completion Verification Design) evaluates built-in self-audit quality per SKILL_STANDARDS.md.

Self-application (T25): Skill-audit's own discovery (this phase) SHOULD use a quick convergence loop: Pass 1 audits all categories, Pass 2 verifies findings are accurate and nothing was missed. Present T20 tally before Phase 3.

Interactive Flow (MUST)

Present ONE category at a time. Wait for user response. Do NOT batch. Even if a rewrite is confirmed at mid-audit, MUST continue one-category-at-a-time interactive flow. The rewrite conclusion does not change the process.

• Correction protocol: If user corrects presentation format, re-present the category in the EXACT original format (full pros, cons, gaps, suggestions with rationale). Never summarize, truncate, or use table-only format.
• Mid-category recovery: If compaction occurs mid-category, re-read state file and restart the current category (partial progress not persisted).
• Cross-category revision: If a later category conflicts with an earlier decision, present the conflict to the user and update state with revision.

Per-Category Procedure (MUST)

1. Present the category — name, purpose, what it evaluates
2. Assess current state — cite specific skill sections/lines
3. List ALL pros — what's working well
4. List ALL cons — what's not working
5. List ALL issues & gaps — what's missing entirely
6. Present suggestions — labeled A, B, C. Recommend with rationale. Two modes: multiple options (with pros/cons per option) or single fix (one recommendation). Use multi-option mode when genuine alternatives exist.
7. Cover EVERY con and gap — each MUST have at least one suggestion
8. Present opportunities section (MUST present, MAY populate) — value-add ideas beyond fixing problems. If genuine ideas exist, present with recommendation + rationale. If none: state "Opportunities: None for this category."
9. Collect decisions — accept/modify/reject/alternatives via conversational Q&A. NEVER use AskUserQuestion — present findings and collect decisions through normal conversation. Delegation: if user says "you decide," accept all recommendations; record delegated-accept.
10. AskUserQuestion check (MUST) — if the audited skill uses AskUserQuestion, flag it as a finding. Skills MUST use conversational Q&A (deep-plan style) instead of AskUserQuestion for all interactive decisions.
11. Tag confidence (SHOULD) — high/medium/low. Low-confidence gets extra confirmation during Phase 4.
12. Save to state file — persist ALL decisions before next category
13. Show progress — "Category 3 of 10 complete. 18 decisions so far."

If >8 suggestions per category, split into sub-batches by theme, severity-first.

Mid-Audit Check (MUST — after Category 5)

"Mid-audit status: Score [N/55], [M] decisions across 5 categories. Top concern: [lowest-scoring category]."

Scope explosion guard: If 3+ categories below 4/10: "Multiple categories show fundamental issues. Continue auditing, or pivot to /skill-creator?"

After Category 12 (MUST)

"Are there quality concerns the 11 categories didn't surface?" Add user findings as additional decisions.

Phase 2a Completion (MUST — single mode only)

See REFERENCE.md for the completion summary template.

Present aggregate summary: all decisions by category, overall score, top findings. Then: "Proceed to implementation with these N decisions? [Y/modify/n]"

Update state file.

---

Phase 2b: Category Audit — batch/multi modes (Batched, MUST)

Mode scope: This phase applies when mode=batch or mode=multi. Line 118 MUST rule (single-category gating) does NOT apply here — it is scoped to Phase 2a only.

Findings-Only Flow (MUST)

Produce findings for ALL 12 categories before collecting any decisions. Decision collection happens in Phase 2.B. This decouples:

• Findings production (this phase — batched)
• User decision gate (Phase 2.B — still interactive per-category)

Per-Skill Procedure (MUST)

For each skill in the batch (one skill if mode=batch, list if mode=multi):

1. For each category (1-12) — follow the Phase 2a per-category procedure steps 1-8 (assess, list pros, cons, gaps, suggestions with recommendations, opportunities) — SKIP step 9 (collect decisions). Findings go to state under findings_by_category.<cat_key>.
2. Save state after each category's findings (MUST — matches Phase 2a save cadence). See REFERENCE.md §Batch Findings Production Procedure.
3. Render markdown to .claude/tmp/skill-audit-<name>-findings.md after all 12 categories complete. See REFERENCE.md §Batch Findings Rendering.

Faithfulness Guarantee (MUST)

Findings produced in Phase 2b MUST be equivalent to what Phase 2a would produce on the same skill — same 12 categories, same pros/cons/gaps/suggestions depth, same REFERENCE.md rubric. Only delivery differs (batched vs gated).

Done when: All skills in the batch have complete state.findings_by_category and rendered tmp file. State saved after each category per skill.

---

Phase 2.A: Cross-Skill Pattern Detection (MUST — multi mode only)

Mode scope: This phase applies ONLY when mode=multi. Skip for batch or single modes.

Purpose

Surface systemic patterns (same gap type, same missing section, same anti-pattern) that appear across multiple skills in the batch — before any decisions are collected. Lets cross-skill context inform per-skill decisions.

Procedure (MUST)

1. After all skills' Phase 2b findings produced
2. Analyze findings_by_category across all skills in the batch
3. Identify patterns appearing in 3+ skills (threshold): same gap label, same missing section, same anti-pattern type
4. Write to parent batch state file cross_skill_patterns field (see REFERENCE.md §Parent Batch State Schema)
5. Present to user before Phase 2.B begins: "Patterns detected across 3+ skills: [list]. These will be noted in per-skill decision prompts."

Done when: Cross-skill patterns written to parent state and presented to user.

---

Phase 2.B: Decision Collection (MUST — batch/multi modes)

Mode scope: This phase applies when mode=batch or mode=multi.

Purpose

Collect accept/modify/reject/alternative decisions for the already-produced findings. Faithful to Phase 2a's per-category decision schema — just decoupled from findings production.

Procedure (MUST)

For each skill in the batch (if multi, iterate through all skills):

1. Reference the rendered tmp findings file (.claude/tmp/skill-audit-<name>-findings.md)
2. For each category (1-12):
   • Present that category's suggestions (read from state, NOT re-analyze — findings are locked from Phase 2b per faithfulness guarantee)
   • For mode=multi: surface any cross_skill_patterns entries affecting this category inline ("also appears in 3 other skills in this batch")
   • Collect decisions conversationally (accept/modify/reject/alternative) — NEVER use AskUserQuestion. Same schema as Phase 2a step 9.
   • Real-time conflict check (MUST): compare new decision vs all earlier decisions in this audit. If conflict detected, present and resolve before continuing.
   • Save decision to state file (MUST — per category)
3. After all 12 categories decided: final sweep conflict pass (MUST backstop) over the full decision set.

Done when: All categories across all skills in the batch have decisions recorded; both real-time and final-sweep conflict checks complete.

---

Phase 2.5: Operational Dependency Check (SA-1, MUST)

Run these checks for every skill with scripts, hooks, or data file dependencies:

Check	Level	What to verify
A. Scripts	MUST	All invoked scripts exist, run without error, produce expected output
B. Hooks	SHOULD	All hooks that feed data are configured and write output
C. Data files	MUST	Every file read has a writer, every file written has a reader
C2. Contracts	MUST (if consumers)	Output artifacts match downstream consumer expectations (format, schema)
D. npm scripts	SHOULD	All npm run commands exist in package.json
E. Docs	SHOULD	All referenced docs/anchors still exist
F. Functions	MAY (MUST if >3)	Internal functions produce output, handle missing input
G. State files	MUST	Schema matches read/write usage, path writable

MUST investigate root causes (SA-3) — don't just note "file missing." Trace the writer, identify WHY it's missing (dead code? never wired? wrong path?).

MUST functionally validate scripts (SA-5) — existence is the floor, not the ceiling. Run each script in dry-run/check mode and verify it produces expected output. A script that exists but fails silently is worse than a missing script.

Present findings with severity and fix recommendations. User decides per finding. Pause for user confirmation before proceeding to Phase 3.

---

Phase 3: Crosscheck + Ecosystem Impact (MUST)

Mode branching:

• mode=single or mode=batch: run this phase per-skill (single skill in scope)
• mode=multi: run this phase ONCE across the batch (see Batched Phase 3 below)

Standard Phase 3 (single / batch modes)

1. Review skill-creator — does it guide creators to avoid the gaps found?
2. Self-audit crosscheck (MUST) — does skill-creator's discovery include self-audit design questions? Does the content checklist require a self-audit phase per SKILL_STANDARDS.md?
3. Present crosscheck summary: gap count + recommendations
4. Adjacent skill contracts (SA-4, MUST) — referenced skills exist, interfaces match assumptions, handoff protocol consistent on both sides
5. Ecosystem impact (MUST) — identify downstream skills/files affected. For each impact, offer actionable solutions (not just notifications). User may address downstream impacts within this audit or defer.

Batched Phase 3 (multi mode only, MUST)

Run the crosscheck ONCE across the batch, not N times:

1. Skill-creator crosscheck — review ONCE against composite gap list (union of skill-creator-affecting gaps across all audited skills)
2. Self-audit crosscheck — ONCE for the composite
3. Adjacent skill contracts — aggregate across batch; identify cross-skill contract inconsistencies (e.g., two audited skills reference each other with mismatched handoff assumptions)
4. Ecosystem impact — aggregate downstream skills/files across all audited skills; dedupe overlapping impacts
5. Present a single batched crosscheck summary with per-skill breakdown + composite findings

Rationale: skill-creator crosscheck converges quickly across skills (gap set is small); N-times is waste. Batched aggregation catches cross-skill inconsistencies that per-skill would miss.

Update state file (parent batch state for multi, per-skill state for single/batch).

---

Phase 4: Implementation (MUST)

1. Priority order (MUST) — highest-impact decisions first
2. Batch related changes (SHOULD) — group edits to the same file
3. Flag conflicts (MUST) — if two decisions conflict, ask user first
4. Low-confidence confirm (MUST) — decisions tagged low get extra confirmation before applying
5. Correction protocol (MUST) — if user corrects presentation format, re-present in the EXACT original format. Never summarize or truncate.
6. Update SKILL.md — keep under 300 lines (MUST)
7. Extract to REFERENCE.md if needed (SHOULD)
8. Update companion files as needed (SHOULD)
9. Apply skill-creator updates if crosscheck found gaps (SHOULD)
10. Run npm run skills:validate (MUST)

For multi-file skills: Audit root SKILL.md, note sub-command coverage gaps. Do not run separate audits per sub-command.

Done when: All accepted decisions applied, skills:validate passes. Update state file.

---

Phase 5: Self-Audit (MUST — before completing)

Read REFERENCE.md for detailed self-audit criteria and report format. Pattern reference: .claude/skills/_shared/SELF_AUDIT_PATTERN.md.

5.0 Run Self-Audit Script (MUST — first step)

Mode branching:

• mode=single or mode=batch: invoke the per-skill self-audit script once:

  node scripts/skills/skill-audit/self-audit.js --target=<audited-skill-name>
  

• mode=multi: invoke in parallel for each skill in the batch. Aggregate ---SUMMARY--- JSON blocks into a composite result; present to user as a batch table (one row per skill) with overall status + counts of must_failed / should_warned.

Parse the ---SUMMARY--- JSON block. Branch:

• overall == "FAIL": present each must_failed dimension to the user with remediation options. Re-enter Phase 4 (Implementation), fix, re-run the script. Per SKILL_STANDARDS.md §Self-Audit ordering: "If self-audit finds failures, re-enter Build, fix, then re-run Self-Audit." For multi mode: failed skills get per-skill remediation; passing skills remain clean.
• overall == "PASS" with should_warned non-empty: present warnings (typically regression no-history note on first runs). User decides acknowledge / fix / defer. Proceed only after explicit decision.
• overall == "PASS" clean: proceed to 5.1 prose verification.

The script covers all 9 MUST dimensions mechanically (Dim 6 is a deterministic cross-reference integrity check, replacing the previous agent-based layer per Session #281 D11). Steps 5.1-5.5 below handle process-compliance + decision-by- decision evidence walkthrough — no LLM agent re-interpretation, deterministic only.

5.1 Re-read All Modified Files (MUST)

Re-read every file modified or created in Phase 4. Do NOT rely on memory.

5.2 Evidence-Based Decision Verification (MUST)

Logging a decision as "PASS" does NOT mean it was implemented. Each decision MUST be verified with objective evidence. See REFERENCE.md for the deterministic verification methods and evidence format.

Deterministic only (Session #281 change): Verification uses grep + diff mechanical checks. No LLM agent re-interpretation. Rationale: another LLM reading the same inputs is echo, not independent verification. Deterministic checks (grep + diff + schema validation) catch the "I thought I wrote it but didn't" failure mode without drift risk.

Two verification layers, ALL required:

1. Grep-based proof (MUST): For each decision, grep the output file for a keyword or pattern that proves implementation. Cite the grep result. If grep finds nothing, the decision is MISSING — not PASS.
2. Diff-based mapping (MUST): Generate git diff of all modified files. Map each decision to the specific diff hunk that implements it. Decisions with no corresponding diff hunk are MISSING.

Present results as decision-by-decision table. If >20 decisions: group by category, expand only PARTIAL/MISSING items; PASS items get one-line summary with grep evidence.

5.3 Process Compliance Verification (MUST)

Verify the skill-audit process was followed. See REFERENCE.md checklist.

5.4 Structural Validation (MUST)

Run npm run skills:validate, check line count, verify cross-references.

5.5 Self-Audit Report → Completion Summary (MUST)

Report is ALWAYS shown first. Completion Summary ONLY after report passes clean. They are sequential, not alternatives. Fix any PARTIAL/MISSING before summary.

Update state file to status: complete.

---

Phase 6: Learning Loop + Closure (MUST)

Auto-learnings (MUST): Generate 2-3 data-driven insights from audit results (lowest-scoring category, most common gap type, recurring patterns across repeat audits). Save to state file learnings field.

Optional user feedback (SHOULD): "Any additional observations?" Accept empty / "none" to proceed. If provided, save to process_feedback field.

On next startup (MUST): If previous audit state exists for the same skill, surface auto-learnings and user feedback from the previous run.

Invocation tracking (MUST):

Context MUST include {"target":"SKILL_NAME","decisions":N} — substitute the actual skill name and decision count. Empty context breaks audit tracking.

# single / batch mode:
cd scripts/reviews && npx tsx write-invocation.ts --data '{"skill":"skill-audit","type":"skill","success":true,"schema_version":1,"completeness":"stub","origin":{"type":"manual"},"context":{"target":"SKILL_NAME","decisions":N,"score":SCORE,"mode":"single"}}'

# multi mode — includes batch_id + skills_in_batch:
cd scripts/reviews && npx tsx write-invocation.ts --data '{"skill":"skill-audit","type":"skill","success":true,"schema_version":1,"completeness":"stub","origin":{"type":"manual"},"context":{"mode":"multi","batch_id":"BATCH_ID","skills_in_batch":["a","b","c"],"decisions":N}}'

Closure summary:

Skill Audit Complete: [skill-name]
Categories: 10 | Decisions: [N] ([M] accepted, [K] rejected)
Overall Score: [N/100] → post-fix: [N/100]
[If repeat: Previous: [N] → Current: [M] | Improved: [list] | Regressed: [list]]
Files modified: [list] | Skill-creator gaps: [N]

---

Guard Rails

• Target doesn't exist: Report error, suggest /skill-creator
• Scope explosion: 3+ categories below 4/10 at midpoint → offer pivot
• Pause/resume: User says "pause" → save state, print progress, exit. Resume: /skill-audit <name> reads state, skips completed categories.
• Multi-file skills: Audit root SKILL.md only; note sub-command gaps
• Phase entry gates: Phase 2 requires warm-up + state init. Phase 3 requires all 10 categories. Phase 4 requires crosscheck + user approval.

---

Compaction Resilience

• State file: .claude/state/task-skill-audit-{skill-name}.state.json
• Update frequency: After every category AND every phase boundary
• Recovery: On resume, read state file, skip completed categories/phases
• Retention: File retained after completion as audit decision record
• State schema: See REFERENCE.md for full schema

---

Integration

• Neighbors: /skill-creator (create/update), /skill-ecosystem-audit (ecosystem-wide), npm run skills:validate (structural)
• Input: Target skill name or path
• Output: Updated skill files + decision record in state file (the state file IS the persistent decision record)
• Handoff: Use /skill-creator for major rewrites after audit
• Multi-skill batches: Use mode=multi (selected in Phase 1.0) to audit 3+ skills in one coherent run. The skill handles batch orchestration natively via Shape Y (audit-all → decide-all → implement-all) with cross-skill pattern detection in Phase 2.A. No team dispatch required — faithfulness to the 12-category rubric is maintained by running the real skill process, not agent simulation.

---

Version History

Version	Date	Description
4.0	2026-04-14	Add batch + multi modes (D7); split Phase 2 into 2a/2b (D8); Phase 2.B decision collection (D9); Phase 2.A cross-skill pattern detection (D20); batched Phase 3 crosscheck (D20); remove code-reviewer agent layer from all modes (D11); parallel Phase 5 self-audit.js dispatch for multi (D12); decouple from audit-review-team (D19). Per skill-audit-batch-mode plan.
3.6	2026-04-14	Phase 5.0: invoke scripts/skills/skill-audit/self-audit.js (pattern-based)
3.5	2026-04-04	Add Category 12 (Completion Verification Design), expand Phase 2.5+3
3.4	2026-03-19	Fix invocation tracking: context MUST include target+decisions+score
3.3	2026-03-15	Add Category 11 (T25 convergence loop), self-application in Phase 2
3.2	2026-03-07	Evidence-based self-audit: grep proof, agent verification, diff mapping
3.1	2026-03-07	SA-1,3,4: Phase 2.5 operational deps, root cause, adjacent contracts
3.0	2026-03-06	Self-audit: 42 changes, routing, guard rails, UX, confidence.
2.0	2026-03-06	Add Phase 5 self-audit. Source: pr-retro audit session.
1.0	2026-03-01	Initial implementation from deep-plan audit of deep-plan