Validate Agent Skills

Version updated for https://github.com/mohitagw15856/pm-claude-skills to version v76.2.1.

  • This action is used across all versions by 0 repositories.

Action Type

This is a Composite action.

Go to the GitHub Marketplace to find the latest changes.

Action Summary

PM Skills is an open-source library of 1099 Agent Skills that provide plain-markdown SKILL.md files to teach AI assistants professional tasks. Each skill bundles the framework, output template, quality checks, and anti-patterns, making it easier for AI assistants to perform tasks accurately at senior professional levels. PM Skills supports Claude, ChatGPT, Gemini, Cursor, Codex, and Hermes Agent, with ready-to-paste exports for other tools.

What’s Changed

The SKILL.md validator that keeps these 1,099 skills honest now runs against yours — as a GitHub Action.

- uses: mohitagw15856/pm-claude-skills@v76
  with:
    path: .claude/skills   # optional — it finds them otherwise

Or without CI: npx pm-claude-skills skillcheck

What it checks

Skill authoring is new enough that the mistakes are the same handful, and every one is cheap to catch:

  • Frontmatter — present, well-formed, and a name that matches its folder
  • The Use when … trigger clause — the highest-value line in a skill, because it’s what a model matches on when deciding whether the skill applies at all
  • Leftover template text — your-skill-name, [Instructions for Claude to follow
  • Structure — a # Title, Quality Checks, Anti-Patterns
  • Length — too short to trigger on, or too long for the trigger budget

Errors fail the build. Warnings are style advice and don’t, unless you pass --strict.

Findings land on the line

The reason to use an action rather than a shell step is annotation. Each finding carries a real line number and appears inline on the pull request diff:

::error file=skills/name-mismatch/SKILL.md,line=2,title=SkillCheck::Frontmatter name "totally-different" does not match folder "name-mismatch".

errors, warnings and skills are exposed as step outputs, populated even when the check fails — so you can comment on a PR or gate a later step on the numbers.

One implementation, including for this repo

scripts/skillcheck.mjs is now a thin wrapper around the published bin/skillcheck.mjs, so this library is validated by exactly the same rules yours is. A linter that exempts itself isn’t worth much.

Also in this release

  • design-system-generate — skill #1099, for the gap between audit the system we have and extract the brand we have: there is no system and something ships on Thursday.
  • /design-review — a workflow recipe that critiques first, then measures. It closes on the split that makes findings schedulable: “nine are a find-and-replace, three need a design decision”.
  • The design skills compute their contrast numbers instead of estimating them. #777777 on white is 4.478 and fails AA; #767676 is 4.542 and passes, and no amount of looking at a screenshot separates those. Six skills plus brand-guidelines call notugly; the MCP server exposes check_contrast as an eighth tool.
  • Generated persona faces on the playground chips and the starter-pack banner.

Fixed since v76.2.0

  • The Marketplace caps an action description at 125 characters and this one was 152. The folded YAML form counts the joined string, so wrapping it changed nothing — it had to actually be shorter.
  • npm-publish now skips a version already on npm instead of failing with E403, which turned the last release red for no good reason.

Zero dependencies. No Docker image. No model call. MIT.