AGENT SKILL

skill-crucible

Audit any agent skill before trusting it — produces a 0-100 quality and security score for SKILL.md folders. Use when asked to check, vet, review, score, or audit a skill, or before installing third-party skills from a marketplace or GitHub.

License 许可
MIT
最近更新
2026-10-17
Stars
★ 1
Forks
0

01 前言 Front Matter

name
skill-crucible
license
MIT
description
Audit any agent skill before trusting it — produces a 0-100 quality and security score for SKILL.md folders. Use when asked to check, vet, review, score, or audit a skill, or before installing third-party skills from a marketplace or GitHub.

02 资源文件 34

03 SKILL.md 正文

skill-crucible

You are auditing an agent skill. Produce evidence, not vibes.

When this applies

  • The user asks to check, vet, score, review, or audit an agent skill
  • A third-party skill is about to be installed from a marketplace, GitHub, or a shared folder
  • The user wants CI gating or a README badge for their own skills

Procedure

  1. Run the scan:

    npx skill-crucible check <path-to-skill-folder>
    

    Useful flags:

    • --online — verify every npx/pip install package name against the real registry (catches hallucinated package names an attacker could claim)
    • --format json -o report.json — machine-readable report
    • --fail-on high — exit 1 at the chosen severity, for CI gates
  2. Interpret the score:

    • 90+ (A) and verdict ship — skill is structurally sound, no findings worth blocking
    • 75–89 (B) — ship after reviewing medium findings
    • 50–74 (C) — revise: weak description, oversized body, or medium-risk patterns
    • below 50 (F) or verdict reject — do not install or publish yet
  3. The safety gate: any critical finding caps the score at 40 and forces reject. Criticals include prompt-override phrasing, download-piped-to-shell, destructive filesystem commands, hardcoded secrets, and package names that do not exist on the registry. Treat a critical as a hard stop.

  4. Report to the user:

    • score, grade, and verdict
    • every critical and high finding with file and line
    • for criticals: recommend not installing the skill until fixed
  5. To publish a badge for a skill the user owns:

    npx skill-crucible check . --format json -o report.json --badge badge.svg
    

    Then reference badge.svg from the repository README.

Limitations to state honestly

This is fast static analysis, not a sandbox. It never executes what it scans. False positives are possible; novel obfuscation can pass. For deep inspection, pair it with a dedicated scanner such as NVIDIA SkillSpector.

04 同仓库更多 Skills