skill-check

skill-check

Linter for agent skill files — validates SKILL.md files against the spec with extensible custom rules.

Regenerate with pnpm run demo:readme (source: scripts/readme-demo.tape).

Install

npx skill-check check .

Global install via curl:

curl -fsSL https://raw.githubusercontent.com/thedaviddias/skill-check/main/scripts/install.sh | bash

Install via Homebrew:

brew tap thedaviddias/skill-check https://github.com/thedaviddias/skill-check
brew install skill-check

Commands

Command	Description
skill-check check [path]	Run validation (and optional security scan)
skill-check new <name>	Scaffold a new skill directory with SKILL.md template
skill-check watch [path]	Watch for changes and re-run validation on save
skill-check diff <a> <b>	Compare diagnostics between two skill directories
skill-check report [path]	Generate a markdown health report
skill-check rules [id]	List all rules, or show detail for a specific rule
skill-check security-scan [path]	Run security scan via agent-scan (mcp-scan)
skill-check init	Create skill-check.config.json template

check options

Flag	Description
--fix	Apply safe automatic fixes for supported findings
--fix --interactive	Prompt before applying each fix (TTY only)
--baseline <path>	Compare against a previous JSON run and show new/fixed counts
--format <fmt>	Output format (see below)
--no-open	Skip auto-opening HTML reports
--no-security-scan	Skip the security scan
--strict	Treat warnings as errors
--lenient	Relax selected strict rules
--fail-on-warning	Exit non-zero when warnings exist

Output Formats

Format	Description
text	Colorized terminal output with ASCII tables, severity badges, and quality scores (default)
json	Machine-readable output including quality scores and optional baseline diff
sarif	SARIF format for security tooling and GitHub Code Scanning
html	Self-contained HTML report with scores, filtering, and dark mode
github	::error / ::warning annotations for GitHub Actions

HTML reports are written to skill-check-report.html (or output.reportPath). In an interactive terminal the report opens in your browser automatically; use --no-open to skip.

View locally: npx skill-check check . --format html or open the file directly: open skill-check-report.html (macOS).

The text formatter includes quality score bars per skill, colorized severity badges, and boxed summaries. An ASCII CLI banner is shown in interactive text mode; set SKILL_CHECK_NO_BANNER=1 to disable it.

Quality Scores

Every check run computes a quality score (0-100) per skill based on five weighted categories:

Category	Weight	What it measures
Frontmatter	30%	Required fields, naming, ordering
Description	30%	Length, "Use when" phrasing
Body	20%	Line/token limits
Links	10%	Broken local and reference links
File	10%	Trailing newlines, formatting

Scores appear in text, html, and json output.

Duplicate Detection

When multiple skills share the same name or identical description, check emits duplicates.name / duplicates.description warnings so agents can reliably differentiate skills.

Baseline Comparison

Save a JSON run as a baseline and compare later:

npx skill-check check . --format json --no-security-scan > baseline.json
# ... make changes ...
npx skill-check check . --baseline baseline.json --no-security-scan

Output shows how many diagnostics are new, fixed, or unchanged.

Quick Start

pnpm install
pnpm run check
pnpm run report

Generate real CLI outputs from multi-skill fixtures:

pnpm run smoke:cli

Smoke output files are written to reports/smoke/. Smoke includes a real security scan run by default. Smoke also includes a real --fix run on a temp copy of the failing mixed fixture. You can control smoke output colors with SMOKE_COLOR=always|auto|never. You can set SMOKE_SECURITY_SCAN=0 to skip security smoke and SMOKE_SECURITY_SCAN_RUNNER=auto|local|uvx|pipx to choose the runner (default: pipx). Use SMOKE_SECURITY_SCAN_SKILLS=/path/to/skills to override the skills path scanned in smoke mode.

Create config with guided setup:

npx skill-check init --interactive

Security Scan

skill-check can validate repos or direct skills directories:

npx skill-check check /path/to/repo
npx skill-check check ~/.claude/skills

check runs the security scan by default. If dependencies are missing, skill-check asks before installing in interactive terminals. In non-interactive/CI environments, use --allow-installs to permit automatic installs.

Run security scan without UV by forcing pipx:

npx skill-check security-scan . --security-scan-runner pipx --allow-installs

Run validation + security scan in one pipeline step with explicit runner:

npx skill-check check . --security-scan-runner pipx

Skip security scan for local/offline linting:

npx skill-check check . --no-security-scan

Apply safe auto-fixes and then re-run validation:

npx skill-check check . --fix --no-security-scan

Interactively choose which fixes to apply:

npx skill-check check . --fix --interactive --no-security-scan

Scaffold a new skill:

npx skill-check new my-skill

Watch for changes during development:

npx skill-check watch . --no-security-scan

Compare two skill directories:

npx skill-check diff skills/ other-skills/

Use GitHub annotations in CI:

npx skill-check check . --format github --no-security-scan

Hard-block dependency installs:

npx skill-check check . --no-installs

Auto-fix Coverage

--fix currently handles deterministic formatting/metadata issues:

• frontmatter.required
• frontmatter.name_required
• frontmatter.description_required
• frontmatter.name_matches_directory
• frontmatter.name_slug_format
• frontmatter.field_order
• description.use_when_phrase
• description.min_recommended_length
• file.trailing_newline_single

Rules requiring human intent (content quality, max-length trimming, broken links, or oversized bodies) remain manual and are reported after fixes are applied.

Use --fix --interactive for per-diagnostic approval prompts (requires TTY).

GitHub Action

Use skill-check directly in workflows:

Marketplace status: not listed in GitHub Marketplace yet. Supported usage today is direct repo tags (uses: thedaviddias/skill-check@v1 or @v1.x.y). See docs/github-action-publishing.md for the publication playbook.

name: skill-check

on:
  pull_request:
  push:
    branches: [main]

jobs:
  validate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: thedaviddias/skill-check@v1
        with:
          path: .

Use --format github for inline annotations on PRs:

      - run: npx skill-check check . --format github --no-security-scan

Enable security scan explicitly (default is disabled in the action):

name: skill-check-security

on:
  pull_request:

jobs:
  validate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: thedaviddias/skill-check@v1
        with:
          path: .
          security-scan: "true"
          security-scan-install-policy: allow
          security-scan-runner: pipx

Emit SARIF and upload to GitHub Code Scanning:

name: skill-check-sarif

on:
  pull_request:
  push:
    branches: [main]

permissions:
  contents: read
  security-events: write

jobs:
  validate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - id: skillcheck
        uses: thedaviddias/skill-check@v1
        with:
          path: .
          format: sarif
          sarif-file: reports/skill-check.sarif.json
      - uses: github/codeql-action/upload-sarif@v3
        with:
          sarif_file: ${{ steps.skillcheck.outputs.sarif-file }}

The action outputs:

• exit-code: CLI exit code
• sarif-file: absolute path to SARIF file when format=sarif
• command: full command used for execution

format=sarif cannot be combined with security-scan=true in the action because security scan output is not SARIF.

Action Release Checklist

• Ensure main contains the desired action.yml and github-action/index.js.
• Create and push an immutable version tag (example: v1.2.0).
• Move the major tag to the latest stable release (v1 -> v1.2.0 commit).
• Verify a workflow using uses: thedaviddias/skill-check@v1 resolves the updated action.

Rules Reference

Run skill-check rules to see all built-in rules with severity and fixable status. Run skill-check rules <id> for detail on a specific rule.

Rule	Default	Fixable
frontmatter.required	error	yes
frontmatter.name_required	error	yes
frontmatter.description_required	error	yes
frontmatter.name_matches_directory	error	yes
frontmatter.name_slug_format	error	yes
frontmatter.field_order	error	yes
description.max_length	error	no
description.use_when_phrase	warn	yes
description.min_recommended_length	warn	yes
body.max_lines	error	no
body.max_tokens	warn	no
file.trailing_newline_single	warn	yes
links.local_markdown_resolves	warn	no
links.references_resolve	warn	no
duplicates.name	warn	no
duplicates.description	warn	no

All rules emit actionable suggestion text to guide fixes.

Releasing

Releases are automated with semantic-release. Pushing to main (after CI passes) runs the release workflow: commits are analyzed for Conventional Commits (fix:, feat:, BREAKING CHANGE:), the version is bumped, CHANGELOG.md is updated, the package is published to npm, and a GitHub release is created.

• Commit messages are validated locally by commitlint (enforced by the commit-msg hook). Use fix:, feat:, docs:, chore:, etc.
• Secrets: In GitHub, set the NPM_TOKEN repository secret (npm automation token with publish scope) so the workflow can publish to npm. GITHUB_TOKEN is provided automatically.

To simulate a release locally (without publishing): pnpm run release:dry-run. It will fail verifyConditions without NPM_TOKEN and GITHUB_TOKEN; in CI both are set.

Docs

• docs/quickstart.md
• docs/github-action-publishing.md
• docs/config.md
• docs/rules.md
• docs/plugins.md
• docs/migration-from-agent-forge.md