Vendor-neutral CLI for validating, linting, and improving portable Agent Skills that follow the agentskills.io specification.

check-skills is designed to be used inline in LLM coding sessions. After an agent creates or edits a skill, tell it to validate that skill:

Create a portable Agent Skill for <task>, then run:
pnpx check-skills validate <skill-path>

For repositories containing many skills, tell the agent to validate the repo recursively:

Validate all skills in this repo with:
pnpx check-skills validate . --recursive --llm

Recommended inline commands:

# After creating or editing one skill
pnpx check-skills validate ./my-skill --llm

# After editing a skills repository
pnpx check-skills validate . --recursive --llm

# Machine-readable output for automation or agent parsing
pnpx check-skills validate . --recursive --json

# CI/spec-only validation against agentskills.io requirements
pnpx check-skills validate . --recursive --no-quality

Fix all errors before finishing. Treat warnings as quality prompts during authoring; use --no-quality when you only want spec compliance.

Validates skill directories.

Options:

• --recursive — discover skills recursively
• --strict — treat warnings as failures
• --json — machine-readable output
• --llm — concise stable output for coding agents
• --quiet — only show skills with problems
• --format github — emit GitHub workflow annotations
• --agent <name> — run optional adapter checks (codex, claude-code, opencode, cursor, windsurf, pi)
• --no-quality — only run spec compliance checks

Explains a validation rule and suggested fix:

check-skills explain missing-trigger-language
check-skills explain invalid-allowed-tools --json

Generates an <available_skills> prompt block with each skill's name, description, and absolute SKILL.md location.

Summarizes a skills directory, including pass/fail counts, warnings, duplicate names, and asset/reference/script usage.

Plans safe automatic fixes. Dry-run planning is the default; use --write to apply deterministic low-risk edits. Doctor reports include stable fix codes and before/after snippets in --json output. Unsafe or prose-heavy cases are reported as manual suggestions instead of being rewritten.

Creates a minimal skill scaffold:

check-skills init my-skill
check-skills init my-skill --references --scripts --assets

check-skills also exposes a typed ESM API:

import { validate_paths } from 'check-skills';

const report = validate_paths(['.'], { recursive: true });

The CLI entrypoint is available as check-skills/cli for tooling that needs to resolve it explicitly.

Public exports include:

• validate_paths, validate_skill_dir, discover_skill_dirs
• read_skill_document, parse_frontmatter
• doctor_path
• collect_prompt_skills, skills_to_prompt
• get_stats, format_stats
• public result, problem, validation, stats, and prompt types

Errors include missing SKILL.md, invalid YAML frontmatter, missing required fields, invalid names, name/directory mismatch, overlong descriptions, invalid license, invalid compatibility, invalid allowed-tools, invalid metadata, and missing local file references.

Validation results include per-skill stats for line count, body word count, estimated tokens, description length/tokens, code blocks, sections, and long paragraphs. Human output shows those stats when a skill has findings; JSON output always includes them.

Warnings include vague descriptions, list-heavy descriptions, first/second-person wording, missing trigger language, oversized SKILL.md files, too many sections or code blocks, long paragraphs, TODO/template placeholders, low description/body keyword overlap, vendor-specific wording in portable skills, empty resource directories, orphaned reference/script/asset files, unreferenced scripts, non-executable referenced scripts, and missing concrete instructions.

agentskills.io is the portable skill standard. check-skills is an authoring-time CLI that helps people and agents validate skills against that standard without depending on Claude Code, Codex, Cursor, Windsurf, OpenCode, Pi, or any other harness.