A skill for discovering, evaluating, packaging, and managing AI skills from community sources.
The open skills ecosystem is growing fast — and so are the risks. Snyk's 2025 ToxicSkills research found that 36% of publicly available agent skills contain vulnerabilities, with over 1,400 malicious payloads identified across skill registries. Nearly 900 malicious skills were discovered in a single registry alone.
Security scanners like Caterpillar, MCP-Scan, and Cisco's skill-scanner have emerged to detect malware and prompt injection. Marketplaces like SkillsMP and Agent Skills Hub handle discovery. But no tool ties the full lifecycle together: finding skills, evaluating them holistically (not just for security, but for quality, structure, and fit), packaging them for installation, and managing what you've got over time.
That's what skill-scout does. It's not another scanner — it's the evaluation and lifecycle framework that sits between discovery and installation, combining security vetting with quality assessment in a repeatable pipeline.
skill-scout gives your AI assistant a repeatable pipeline for skill acquisition:
- Scout — Search tracked GitHub repositories for new skills, pull candidates into an incoming queue
- Vet — Score each skill on a 25-point checklist covering structure, value, quality, security, and adaptability. Skills with executable code are flagged for human review.
- Adopt or Reject — Move skills to a vetted library or reject with documented reasoning
- Package — Wrap vetted skills into
.skillfiles (the installable format) ready for the platform UI - Load — At the start of a session, identify which vetted skills are relevant and load them into context
- Update — Periodically check if adopted skills have been updated upstream and re-vet changes
- Cull — Retire skills that are obsolete, superseded, or no longer relevant to your work
The vetting checklist (references/vetting.md) implements defense-in-depth:
- Mandatory file audit before scoring — every file in the skill directory is inventoried and inspected, not just SKILL.md
- Quick reject criteria for obfuscated code, base64 content, prompt injection patterns, and credential access
- Human gate for executable content — skills with scripts, hooks, or command-bearing config files cannot be auto-adopted
- Supply chain awareness — popularity is not security; trending skills get more scrutiny, not less
- Trust tiers — platform vendor skills get lighter review, unknown sources get deep vetting
Package the skill and install via the Skills UI:
python3 scripts/package_skill.py ./skill-scout/ ./Then upload skill-scout.skill through claude.ai Settings → Skills.
Copy the skill-scout/ directory into your skills folder:
cp -r skill-scout/ ~/.claude/skills/skill-scout/The vetting checklist and inventory script work independently. You don't need to install the full skill to use them:
references/vetting.md— Use as a review checklist for any skill you're consideringscripts/inventory.py— Run against any directory of skills to get a status report
Tell your AI assistant:
- "Scout for new skills" — Searches tracked sources, pulls candidates into incoming/
- "Vet incoming skills" — Scores candidates, adopts or rejects with documentation
- "Package [skill-name]" — Wraps a vetted skill into an installable .skill file
- "Load skills for this task" — Identifies and loads relevant vetted skills
- "Run skill inventory" — Reports current state of the pipeline
- "Check for skill updates" — Compares adopted skills against upstream versions
- "Run cull pass" — Reviews vetted library for skills to retire
skill-scout/
SKILL.md — The meta-skill (instructions for the AI)
scripts/
inventory.py — Skill inventory scanner with executable content detection
package_skill.py — Single and batch .skill packager
references/
vetting.md — 25-point vetting checklist with security framework
sources.md — Source registry template with community examples
Your skill library (created on first use):
skill-library/
incoming/ — Candidates pulled from sources, awaiting review
vetted/ — Reviewed and approved, ready to load and package
rejected/ — Declined with documented reasons
archive/ — Retired skills kept for reference
logs/ — Scout run history and vetting decisions
zips/ — Packaged .skill files ready for installation
# Basic usage
python3 scripts/inventory.py ./skill-library
# With system-installed skills
python3 scripts/inventory.py ./skill-library --system-skills ~/.claude/skillsReports all skills across the pipeline, parses YAML frontmatter, and flags skills containing executable content or suspicious patterns.
# Package a single skill
python3 scripts/package_skill.py ./skill-library/vetted/dev/my-skill/ ./skill-library/zips/
# Package all vetted skills at once
python3 scripts/package_skill.py --all ./skill-library/vetted/ ./skill-library/zips/Produces .skill files (zip format) ready for installation through the platform UI.
skill-scout uses the SKILL.md format documented by Anthropic. The format is plain markdown with YAML frontmatter — portable to any tool that supports markdown-based skill or instruction files.
.skill packaging follows Anthropic's official format from the skill-creator example skill.
skill-scout is designed to complement, not replace, existing tools in the ecosystem:
- Security scanners — Caterpillar, MCP-Scan, Cisco skill-scanner handle deep automated security analysis. Use them alongside skill-scout's vetting checklist for defense in depth.
- Marketplaces — SkillsMP, Agent Skills Hub handle discovery at scale. Use skill-scout to evaluate what you find there before adopting.
- Skill builders — Claude Code Skill Factory, Superpowers help create skills. Use skill-scout to vet the output.
- skill-creator (Anthropic example skill) — Builds and evals new skills. skill-scout's Vet workflow integrates skill-creator's eval loop for complex skill testing.
See CONTRIBUTING.md. Security improvements to the vetting checklist are the highest-value contributions — if you've encountered an attack pattern or evasion technique the checklist misses, please open an issue.
MIT