skill-auditlisted

# Skill Audit **Repair** half of the learn-on-the-job loop. `distill-skill` mints skills; this skill stops them **rotting**. A stale skill is worse than no skill — it auto-loads a wrong map. **Target:** skill name = one skill; empty = all repo-local skills. | Flag | Meaning | |------|---------| | `--report-only` | Never write (default for multi-skill runs) | | `--fix` | Safe mechanical fixes only, after the report | | `--changed-since <ref>` | Only skills whose sources changed since `<ref>` | ## Hard principle **Audit detects; distill re-verifies.** Cheap and non-destructive by default. With `--fix`, only: lint, tighten over-broad descriptions, and user-confirmed PRUNEs. Content refresh → print `distill-skill <name>`, never silent body rewrite. ## Ledger Sidecar: `.agents/distill-skill-ledger.json` and/or `.claude/distill-skill-ledger.json`. Schema: [../distill-skill/references/ledger-schema.md](../distill-skill/references/ledger-schema.md). Missing ledger → infer `source_of_truth` from cited paths; full re-check. Never trust ledger over a live file. ## Procedure ### 1. Inventory Glob `.agents/skills/*/SKILL.md` and `.claude/skills/*/SKILL.md`. Load matching ledgers. User-global skill dirs: read-only, overlap only. Same name in two project trees → overlap flag. ### 2. Lint Broken frontmatter · `name` ≠ directory · generic description · lines >120 warn / >160 split. These are `--fix`-safe. ### 3. Drift (core) Resolve `source_of_truth` (ledger, else body paths)