skill-audit

Run the skill-audit checklist against the skill directory: read every file, score items S1-S4, C1-C6, D1-D2, E1, P1-P3, DP1-DP5, and M1-M4 as a fraction of applicable points, then write a scorecard ranking fixes by impact. Quote the offending line for every failing item.

Run the bundled check_reference_health.py script against the skill directory or the whole tree. It reports dangling pointers that resolve nowhere and, with --orphans, skills where over 40% of reference files are unreachable, exiting 1 on any dangling pointer.

skill-audit scores a skill's structure and design statically against a checklist and outputs fixes, without measuring output quality. skill-improver runs the skill against evaluation cases and hill-climbs on measured results. Audit is the fast first pass before spending tokens on optimization.

Yes. The DP1-DP5 items flag dated prompting such as ambient CRITICAL/NEVER emphasis, step-by-step scaffolds, update suppressors, anti-formatting rules, and numeric output clamps written for retired models. Each finding is tagged High, Medium, or Low confidence with git blame used to date rules.

Because routing text like the frontmatter description may carry calibrated urgency since skills under-trigger, while behavioral text should explain rather than shout. These look identical to a grep, so the audit classifies text by function before flagging emphasis as a DP1 failure.

Split when the file exceeds 500 lines or 20KB with duplicate-topic reference clusters and many orphaned references, which indicates mega-file accretion. The references/mega-file-split.md guide covers cluster mapping, parallel consolidation merges, and rewriting SKILL.md as a router with a grouped reference index.