chenhz01

Skill Scorecard — Testing skill for Claude Code

Testing community

A 30-second safety scorecard for AI agent skills: run it on any skill folder or GitHub repo before installing.

How to install Skill Scorecard

This entry records only its repository, not the path inside it, so there is no exact command to give. Open chenhz01/skill-scorecard and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Skill Scorecard does

A 30-second safety scorecard for AI agent skills: run it on any skill folder or GitHub repo before installing. Seven structural checks (frontmatter, SPEC, license, scripts/ risk, opaque blobs, injection scan, size), A-D grading, zero third-party dependencies.

Alternatives in Testing

Pua — Use when the user invokes /pua or asks for PUA mode, try-harder/retry help, change-approach coaching, completi 19.5k ★
Agent Skills Spec — A skill is a folder of instructions, scripts, and resources that agents can discover and load dynamically to p 1.9k ★
Add Scaffold — 按 14-phase 工作流为指定 app 编写清理 scaffold（含需求采集、实测勘测、TOML、safety test、UI 集成） 1.3k ★

README

skill-scorecard

A 30-second safety scorecard for AI agent skills — run it on any skill folder or GitHub repo before you install it.

**Part of the Evidence-first Agents suite** — tooling that makes AI agents accountable instead of just capable: [answer-contract](https://github.com/chenhz01/answer-contract) (output discipline) · [skill-spec](https://github.com/chenhz01/skill-spec) (spec discipline) · [memory-wiki](https://github.com/chenhz01/memory-wiki) (memory discipline) · [skill-os](https://github.com/chenhz01/skill-os) (the assembly line). Same author, same zero-dependency philosophy.

Why

GitHub's own documentation warns that third-party agent skills may contain prompt injections, hidden instructions, or malicious scripts — and research on marketplace skills found roughly 1 in 8 carries a serious defect, with skills bundling a `scripts/` directory at ~2x the odds. Yet the standard install advice is still "stick to trusted sources", with no tool to check anything in under a minute.

skill-scorecard is that check: seven structural probes, one score, zero dependencies.

Quick start

# a local skill folder
python tools/skill_scorecard.py path/to/skill/

# or a GitHub repo (downloaded to a temp dir, nothing installed)
python tools/skill_scorecard.py https://github.com/owner/some-skill

Output: a grade (A–D), a score out of 100, and the findings behind it. Exit code 1 on any FAIL — wire it into a pre-install hook.

The seven checks

check	what it asks
SKILL.md present	is this a skill folder at all?
YAML frontmatter	does it declare name + description?
SPEC contract	is behavior pinned to a falsifiable spec?
license	are usage terms stated?
scripts/ directory	does it ship code the agent may run? (~2x defect odds — read it first)
hidden content	opaque blobs that could be encoded payloads?
injection scan	known prompt-injection / exfiltration phrasing?