skill-health-check

Run the skill-lint.js script against your skills directory to get a per-skill report covering routing signatures, When-NOT sections, output definitions, and verification checklists. Findings are ranked P0 (must fix), P1 (should fix), and P2 (suggestion) with an overall pass/fail gate.

A good description encodes three routing cues: what the skill does, when to use it, and when not to use it with redirects to sibling skills. The lint check flags descriptions with fewer than two of the three Use/Avoid/Output cues as P1 or P2 findings.

Pass --skills-dir and --agents-dir flags to point at non-default locations, and add --json for machine-readable output or --fix-hint for remediation suggestions. Exit code 0 means all pass, 1 means warnings only, and 2 means errors were found.

Yes, it computes Jaccard similarity between all skill descriptions and flags pairs above 60% similarity as P2 findings. This helps prevent routing ambiguity where two skills compete for the same user triggers.

Do not use it for code review, document review, or creating new skills; those belong to dedicated review or creator skills. It also does not check .claude directory structure, which is handled by a separate claude-health check.