do-skills-auditlisted

# Skills Audit Audits every skill the repo has — `.claude/skills-global/` (synced to all machines) and `.claude/skills/` (project-only) — plus user-level `~/.claude/skills/` orphan detection. Two layers: a **deterministic lint** (script; code checks what code can verify) and an optional **architecture pass** (`--arch`; model judgment against a fixed rubric). ## Why this audit exists (the economics) A skill costs context in three places, each with different economics: 1. **Description** — ships in *every* session, used or not. The fleet total is the scarcest resource here: budget 4,000 chars (~2% of context), per-skill target ≤120. 2. **Body (SKILL.md)** — loads once per invocation. Only what every invocation needs. Cap 500 lines. 3. **Sub-files** — load on demand. Reference tables, edge cases, templates live here. Always-true policy belongs in CLAUDE.md, not in any skill. Most findings are misplacements across these three boundaries. ## Quick start ```bash python .claude/skills-global/do-skills-audit/scripts/audit_skills.py $ARGUMENTS ``` **If `$ARGUMENTS` was not substituted** (the command shows a literal `$ARGUMENTS`): extract whatever followed `/do-skills-audit` in the user's message and pass it through; no flags means default behavior. | Flag | Description | |------|-------------| | `--fix` | Auto-fix trivial issues (missing name, whitespace, untracked build artifacts) | | `--json` | JSON output (contract in the script docstring; consumed by the skills-audit r