Skill Maker

Create and validate Claude Code skills following best practices.

Quick Start

Create a new skill:

1. Choose a template: simple, with-references, or full
2. Create directory: skills/<skill-name>/
3. Write SKILL.md following the template
4. Run validation: skills validate <skill-name>

Templates

See templates/ for complete templates.

Frontmatter Specification

Required fields:

---
name: skill-name          # kebab-case, unique identifier
description: |            # Multi-line recommended
  What the skill does. Use when: (1) condition, (2) condition.
  Specific symptoms, error messages, or file types it handles.
---

Optional fields:

category: testing|development|documentation|refactoring|security|performance
user-invocable: true      # Can be called via /skill-name
disable-model-invocation: false  # Prevent automatic loading
allowed-tools: Read,Write # Comma-separated tool names
context: fork|inline      # How skill is invoked
agent: agent-name         # Specific agent to use

See references/frontmatter-spec.md for details.

Description Quality

Good descriptions enable semantic matching. Include:

1. What it does (1 sentence)
2. Trigger conditions (Use when:, Helps with:)
3. Specific context (error messages, file types, frameworks)

Good Example

description: |
  Fix "ENOENT: no such file or directory" errors in npm monorepos.
  Use when: (1) npm run fails with ENOENT, (2) paths work in root
  but not packages, (3) symlinks cause resolution failures.
  Covers Lerna, Turborepo, and npm workspaces.

Bad Example

description: A skill that helps with npm problems.

See references/description-guide.md.

Quality Checklist

Before finalizing a skill, verify:

• Name is kebab-case and descriptive
• Description > 50 characters
• Description includes trigger conditions
• Category is valid (if specified)
• Referenced files exist (references/, docs/)
• No slop patterns (placeholder content)
• Content is actionable and specific
• Tested with skills validate

See references/quality-checklist.md.

Validation

Run validation on your skill:

# Validate specific skill
skills validate <skill-name>

# Validate all skills in project
skills validate

# JSON output
skills validate --json

Validation checks:

• Frontmatter format and required fields
• Description quality (length, trigger conditions)
• Category validity
• Slop pattern detection
• Reference file existence

Slop Detection

The validator detects common slop patterns:

Skill Chaining

This skill works with:

• claudeception: Extract learnings → create skill with skill-maker
• tdd: Write tests first when adding CLI validation features
• dogfood-skills: Use skills validate after creating skills

Creating Your First Skill

1. Identify the knowledge: What non-obvious solution did you discover?
2. Check existing skills: Is there already a skill for this?
3. Choose template: Start with simple unless you need references
4. Write SKILL.md: Follow frontmatter spec and description guidelines
5. Validate: Run skills validate <name>
6. Test: Use the skill in a real scenario

Directory Structure

All skills live in the root skills/ directory (canonical location):

skills/<skill-name>/
├── SKILL.md              # Required: Main skill file
├── references/           # Optional: Supporting documentation
│   ├── guide.md
│   └── examples.md
├── templates/            # Optional: Code/config templates
├── scripts/              # Optional: Helper scripts
└── .provenance.json      # Optional: Source tracking

The .claude/skills/ symlink makes skills available to Claude Code. Never create skills directly in packages/skills/skills/ (generated at build time).

Tips

• Start simple, add complexity only when needed
• Keep SKILL.md focused on actionable guidance
• Use references/ for detailed documentation
• Include concrete examples, not just theory
• Update skills when you learn more