Skill Builder

Build Skills That Work

Create skills that are token-efficient, well-structured, and actually useful.

Core principle: SKILL.md should be SHORT. Move details to auxiliary files.

References:

• structure.md — Anatomy of a good skill
• patterns.md — Proven patterns with examples
• checklist.md — Pre-publish validation

File Structure

skill-name/
├── SKILL.md      ← Short! (~30-50 lines)
├── [topic].md    ← Details, loaded when needed
└── FILES.txt     ← Lists all files to publish (one per line)

FILES.txt example:

SKILL.md
patterns.md
reference.md

⚠️ Publishing Rules

Publishing is serious. Each publish is permanent and public. No spam, no experiments.

1. Never publish without verification
2. Never change slug after first publish (breaks installations)
3. Always verify name, description, version match your files
4. Always confirm which files will be published

Pre-Publish Verification

Before ANY publish, send verification to user with:

• Slug (exact)
• Name (exact)
• Version
• Description (exact)
• Files list (from FILES.txt)

Recommended: Generate verification programmatically (script/code) to guarantee accuracy. A PDF or structured format that pulls values directly from your files ensures no copy-paste errors.

User can configure verification level:

• Full review (PDF with all content)
• Summary only (metadata + file list)
• Trust mode (auto-publish, not recommended)

What Makes Skills Good

• Concise — Models are smart, don't over-explain
• Progressive disclosure — Details in separate files
• Clear triggers — Description says WHEN to use
• Actionable — Instructions, not explanations

Before publishing: Run checklist.md