English · 中文

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.

• 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.

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🧹 Outdated patterns: 5 entries flagged, see patterns.md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 runtime does another
- [ ] Clean outdated patterns

The dead skill count is not the interesting part. The interesting part is the rule-execution gap: skills your P0 rules require, with zero actual invocations. The gap between what you wrote down and what actually runs is real, and most people don't measure it.

Claude Code:

git clone https://github.com/runesleo/claude-skill-audit ~/.claude/skills/skill-audit

Then trigger: /skill-audit or "run a skill tree health check".

Other AI harnesses:

SKILL.md is plain markdown with bash blocks. Port the logic — there's nothing Claude-specific about counting invocations or grepping for references.

• Claude Code with at least a handful of skills configured
• A PreToolUse hook writing to ~/.claude/skill-usage.log (the audit is only as good as its data source — without auto-instrumentation, you're guessing)
• Standard skill directory at ~/.claude/skills/

That's it. No API keys, no external services, no LLM calls — pure local file inspection.

• No auto-delete — skills are muscle memory. Some "dead" skills are seasonal (a weekly review fires once a week).
• Archive over delete — candidates go to _archived/ with a 30-day window. Easy to revive if you missed something.
• Human in the loop — the audit identifies, you decide. No autonomous cleanup.
• Data foundation first — without the PreToolUse hook writing skill-usage.log, this skill is just printing vibes.

Born from @HiTw93's Waza (MIT). Waza decomposes engineering into 8 dimensions (think / design / hunt / check / read / write / learn / health) and turns each into a runnable Claude skill. When I mapped my 62 skills against those 8 dimensions, the /health slot was empty — nobody was auditing what was used, dead, or conflicting. This skill exists to fill that slot. Without Waza, this skill wouldn't exist.

The horizontal scanning logic (dead candidate detection, conflict diagnostics, expired pattern flagging, report formatting) is local implementation.

Detection

• Orphan reference detection — hook references a skill that no longer exists in ~/.claude/skills/
• Cross-machine audit — reconcile local Mac with remote VPS skill state
• Baseline diff — highlight what changed since the last audit

Reporting

• HTML dashboard with trend visualization over time
• Export to Obsidian Daily Note format
• Slack / Discord webhook for scheduled audits

Leo (@runes_leo) — AI × Crypto independent builder. Trading on Polymarket, building data and trading systems with Claude Code and Codex.

leolabs.me — writing · community · open-source tools · indie projects · all platforms.

X Subscription — paid content weekly, or just buy me a coffee 😁

Learn in public, Build in public.

MIT — fork freely. When you do, keep the credit to @HiTw93's Waza so the chain stays intact.