Skill Auditor — Pre-Deployment Quality, Security & Drift Audit for Claude Skills

Before you ship a Claude Skill to your team, your customers, or a marketplace — audit it. Score its trigger description, its tool scope, its injection surface, and whether its prose still matches what its bundled files actually do.

What this Claude Skill does

• Runs a two-phase audit on any Claude Skill — hand-written or produced by an automated doc/repo-to-skill generator — covering structure, security, and drift
• Scores the skill's trigger description against ambiguity heuristics, so you catch a skill that will silently never fire, or one that fires on everything
• Reviews every instruction for tool-scope creep, injection surface (external content directing agent behavior), and credential-handling risk against a 25-point checklist
• Ships a working, standard-library-only Python script that parses SKILL.md and flags every file reference that doesn't resolve — the single most common defect in shipped skills
• Produces a scored report (0–100 across five categories) with a hard verdict: Ship, Ship with fixes, or Do not ship — every finding traced to an exact line and a concrete fix

Built for

• Anyone building or maintaining a library of Claude Skills who wants a repeatable pre-release check instead of a read-through
• Teams generating skills with an automated documentation or repo-to-skill pipeline who need a judgment pass before trusting the output
• Marketplace sellers and internal platform teams who need to show a skill was checked for prompt-injection and permission-scope risk before distribution

What's inside

• SKILL.md — the full skill file with frontmatter and the two-phase methodology: structural & security audit, then drift & coverage audit
• scripts/audit_skill.py — a standard-library-only Python script that parses a target skill's frontmatter and resolves every file reference in its prose against what's actually on disk
• reference/security-checklist.md — 25 checks across trigger quality, tool-scope discipline, injection resistance, secrets handling, and drift/completeness
• reference/scoring-rubric.md — the 0–100 scoring model and Ship / Ship with fixes / Do not ship thresholds
• templates/audit-report-template.md — the report structure the skill fills in for every audit
• LICENSE.md — usage terms

How a Claude Skill works

This is a Claude Skill — a single SKILL.md file Claude reads as instructions for a specific kind of work. Drop it in your Claude Project, .claude/skills/ folder, or wherever Claude can read it. Claude triggers the skill automatically when your prompt matches its description, then runs the methodology end to end.

Author

Built by Ukiyo Productions — a strategy-led branding and growth agency. Part of a 100+ skill marketplace built for founders, operators, consultants, and agencies.