skill-audit

CLI that audits, validates, and inventories agent skills (SKILL.md files) across any ecosystem

Audit, validate, and inventory agent skills (SKILL.md files) across any agent ecosystem.

The agent skills ecosystem is exploding — Claude Code, Codex CLI, Cursor, Gemini CLI, and the 80K+ star agent-skills repo all use SKILL.md files. But there’s zero tooling to manage them. skill-audit fills that gap.

Features

  • audit — Scan a directory of SKILL.md files and validate format compliance. Detects missing required fields, broken references, duplicate skill names, and overlapping descriptions.
  • inventory — List all installed skills in a tree or JSON output with descriptions, versions, and tags.
  • diff — Compare two skill directories to see what was added, removed, or changed.
  • --json — All commands support machine-readable JSON output for piping into other tools.

Install

pip install skill-audit

Or from source:

git clone https://github.com/vikasudasi/skill-audit.git
cd skill-audit
pip install -e .

Usage

# Validate a directory of skills
skill-audit audit ./skills

# Pretty-printed inventory tree
skill-audit inventory ./skills

# JSON output for scripting
skill-audit inventory ./skills --json

# See what changed between skill sets
skill-audit diff ./old-skills ./new-skills

# Quick help
skill-audit --help

Example

$ skill-audit audit ~/.hermes/skills
Audit result for /root/.hermes/skills: PASS
Skills parsed: 82
Errors: 0
Warnings: 2

Warnings:
- claude-code/SKILL.md:3 [similar_description] Description is very similar to codex
- codex/SKILL.md:3 [similar_description] Description is very similar to claude-code

Exit Codes

Code Meaning
0 All good
1 Validation errors found
2 File/IO error

Validation Rules

  • YAML frontmatter must be between --- delimiters
  • Required fields: name, description
  • version should be semantic versioning if present
  • related_skills and linked_files paths must be valid
  • Duplicate skill names across the directory are flagged
  • Descriptions that are too similar (edit distance ≤ 10) trigger warnings