skill-auditor

Load the skill with skill_view or read_file, then score it across five dimensions: frontmatter and description, exact commands, pitfalls, verification steps, and structure. Sum the points out of 100 and map the total to an A-F letter grade with ranked fix suggestions.

A description fails discovery when it is too vague for the router to match, lacks a clear trigger class, or overlaps with other skills. Hermes only sees the one-line description before deciding whether to load the full SKILL.md, so generic phrasing means the skill never fires.

List registered skills with skills_list, independently inventory every SKILL.md on disk, validate frontmatter and content size, then grade structural issues as review signals. Apply only high-confidence fixes, rerun the validator, and require zero failures before completion.

Yes, inspect the supplied SKILL.md path directly rather than relying on registry discovery. Do not trust a zero exit code from hermes skills inspect alone; require explicit success output, then audit the registered copy after installation with hermes skills audit.

Skills with vague steps, missing exact commands, or absent verification checks cause silent failures on smaller models that cannot infer intent. The audit rubric flags these gaps in the exact commands and verification dimensions so they can be fixed before deployment.