Claude Skill Audit

claude-skill-audit

English · [中文](README.zh-CN.md)

A health check for your Claude Code skill tree. Find what's actually used, what's dead, what's stepping on each other, and which iron-rule skills you wrote but never fired.

What you get

• Inventory — how many skills, hooks, aliases you have
• Usage ranking — real invocation counts (data source: ~/.claude/skill-usage.log, written by a PreToolUse hook)
• Dead skill detection — 30 days of zero triggers + no hook binding + no CLAUDE.md reference → archive candidate
• Rule-execution gap — the killer feature: skills your iron rules say must fire, but never did. Writing rules and not running them is the same as not writing them.
• Conflict diagnostics — skill groups with overlapping trigger words

Doesn't auto-delete. Surfaces candidates, you decide. Suggested workflow: move to `_archived/`, observe for 30 days, then permanently remove.

How it works

You: /skill-audit

AI:
📊 Skill Tree Audit — 2026-04-17

📈 Inventory: 61 skills / 4 hooks / 12 aliases
🔥 Top 5 by usage:
   session-end(262) / leo-style(182) / codex:rescue(119) / codex(112) / humanizer(92)
   — top 5 = 92% of total calls. Long tail of 40+ skills averages <10.

💀 Dead candidates (15):
   auto-dream / content-analyzer / design-review / graphify / mcp-builder /
   notebooklm / obsidian-sync / search-skill / skill-from-github /
   skill-from-masters / subagent-driven-development / tempo-request /
   video-auto / video-distribute / workflow-architect

⚠️ Rule-execution gap (6 skills, 0 calls despite rule references):
   systematic-debugging — P0 rule "trigger on 3 consecutive failures",
                          fired exactly 0 times in 1388 invocations
   planning-with-files / opensource-publish / leo-evolution /
   tg-review / video-editing-coach

🧹 Outdated patterns: 5 entries flagged, see patterns.md

🎯 Action items:
- [ ] Archive 15 dead candidates
- [ ] Investigate the gap — your rules say one thing, your runti