audit-skills

What problem does it solve? As a repository of AI agent skills grows, it accumulates duplicated rules, oversized SKILL.md files, stale TODO markers, and skills lacking self-improvement references. This Skill automates the audit of the entire skill corpus so quality regressions are detected instead of silently compounding. ## Core Features & Use Cases - Corpus-wide scanning: Runs the deterministic bash script tools/audit-skills.sh to collect LOC counts, normalized sha1 hashes, and n-gram similarity, then detects duplicates (Jaccard 5-gram >= 0.85), bloated skills, and stale markers. - Tiered apply modes: Generates a dated audit report with prioritized actions — P0 fixes applied automatically (add tier fields, create empty RUNBOOK/FINDINGS/LESSONS files), P1 with confirmation, P2 only by explicit action id. - Self-improvement command: The fix subcommand records an agent error into NEVER blocks, LESSONS.md, or FINDINGS.md with PROTECTED:RULE markers so the same mistake cannot recur. - Use Case: After a large refactor of your claude-rules repository, run the audit to find duplicated role skills across projects, skills missing a tier field, and any critical PROTECTED:RULE that was silently dropped. ## Quick Start Ask the agent to run an audit of all skills in the claude-rules repository and show the summary report with prioritized actions.