skill-linter

skill-linter checks a repo of Agent Skills before you commit or release it, and keeps the counts and version numbers in its docs current. One [tool.skill-linter] table in pyproject.toml configures it, so a new skill repo gets the same checks as the others without copying validate-skills.py, validate-markdown.py and update-docs.py into it.

It needs Python 3.11 or later. PyYAML is the only dependency, used for workflow YAML.

uvx skill-linter check        # run without installing
pipx install skill-linter     # or install it once

Status: early (0.x). Rule names and config keys may still change between minor versions.

Commands

skill-linter check   [--only GROUP[,GROUP]] [--skill NAME] [--format text|json] [--root PATH]
skill-linter stamp   [--check] [--dry-run] [--root PATH]
skill-linter release [--format text|json] [--root PATH]
skill-linter rules   [--format text|json]

• check runs every enabled rule except the release checks.
• stamp counts skills, references, workflows and domains, then writes those counts and the version from version.json into the files you list. --check writes nothing and exits 1 when anything is stale.
• release checks the tree the plugin directory will read (plugin_root or a ship list), then runs check.
• rules lists every rule with its group, layer and severity after your overrides.

skill-linter --version prints the installed version. Exit codes: 0 when there are no errors, 1 when there are errors or stale stamps, 2 for a usage or config error. python -m skill_linter runs the same CLI.

Rules

Rules come in three layers:

1. The Agent Skills specification (SPEC-01 to SPEC-10). You can't switch these off.
2. Conventions shared across the skill repos: frontmatter metadata, the description rule, Markdown lint, workflow definitions, hidden conventions such as the docs backlink and title safety, and the plugin-directory release checks (CONV, DESC, MD, WF, H, P). Each one can be switched off or lowered to a warning.
3. Per-repo policy (POL, LINK, and your own X- rules). A policy rule runs only once you give it a value.

Every rule has an ID and a name, such as POL-03 and section-order. Output shows both, and the config accepts either. Run skill-linter rules for the full list.

Cross-skill references

A skill can cite a file in a sibling skill or in a shared folder, written from the skill root: `../vue-expert/references/components.md` or `../_shared/markup-convention.md`. The spec wants every reference to stay inside the skill folder, so SPEC-10 (reference-depth) leaves these paths to two other rules:

• CONV-15 (cross-skill-references) checks that the file exists, that the folder is a skill or is listed in shared_dirs, and that a cited skill appears in metadata.related-skills (a warning when it doesn't).
• P17 (cross-skill-targets-ship) checks that the cited file ships with the plugin. A skill installed on its own won't find its sibling, so these references only work when the whole plugin is installed.

Configuration

Everything lives under [tool.skill-linter] in the repo's pyproject.toml. Unknown keys and unknown rule IDs are errors, so a typo can't switch a check off without anyone noticing. tests/adoption/ holds complete configs for claude-skills and writing-with-agents.

[tool.skill-linter]
skills_dir = "plugin/skills"
shared_dirs = ["_shared"]
commands_dir = "plugin/commands"
plugin_root = "plugin"
docs_url = "https://example.github.io/my-skills/skills/{domain}/{name}/"

[tool.skill-linter.policy]
workflow_steps = [5, 7]
license = "MIT"

[tool.skill-linter.severity]
plugin-icon = "off"   # or P10

[[tool.skill-linter.rules.custom]]
id = "X-lead-support"
target = "body"
pattern = '(?is)\blead\b.*\bsupport\b'
severity = "warning"
message = "Name a lead and a support role"

Custom rules are regular expressions applied to a frontmatter field, a section, the last line, the body or the reference files. Config never runs code.

Stamping

Each [[tool.skill-linter.stamp.files]] entry names a file and how to update it:

[[tool.skill-linter.stamp.files]]
path = "README.md"   # .md, .mdx and .html files get marker stamping by default
patterns = [
  { key = "version", regex = 'version-([\d.]+)-blue\.svg' },
  { key = "referenceFileCount.{skill}", regex = '(?m)^\s+(?P<skill>[\w-]+)/\s+\(SKILL\.md \+ (?P<count>\d+) reference(?P<s>s?)\)' },
]

[[tool.skill-linter.stamp.files]]
path = ".claude-plugin/marketplace.json"
json = { "/metadata/version" = "version", "/plugins/0/version" = "version" }
patterns = [{ key = "skillCount", regex = '(\d+) specialized skills' }]

• Markers: <!-- SKILL_COUNT -->67<!-- /SKILL_COUNT -->, and likewise WORKFLOW_COUNT, REFERENCE_COUNT, DOMAIN_COUNT and VERSION. SKILL_COUNT:<domain> counts one domain and REFERENCE_COUNT:<skill> counts one skill's references.
• json maps a JSON Pointer to a key and replaces only that value, leaving the file's formatting alone.
• patterns replace the one group in each regex. A pattern that matches nothing is an error, which catches a sentence reworded after its pattern was written.
• The key referenceFileCount.{skill} stamps a count per skill, for places where an HTML comment would show up as text, such as a file tree inside a code block. Its regex names the skill and count groups. The optional s group becomes s unless the count is 1, and a line that names no skill is an error.

Development

uv sync
uv run pytest
uv run ruff check . && uv run ruff format --check .

The adoption tests export pinned commits of the two skill repos with git archive and skip when those checkouts aren't present. One of them runs each repo's own update-docs.py and skill-linter stamp on the same change and checks that every file comes out identical.