Skill Creator

Create modular, token-efficient skills that extend Claude's capabilities.

Critical: Description = Trigger

The description field in frontmatter is how Claude decides when to load a skill.

Pattern:

---
name: pdf-toolkit
# GOOD - trigger info in description
description: "Use when processing PDFs, extracting forms, merging documents, or filling PDF forms. Load for any PDF manipulation task."
keywords: pdf, extract, merge, forms
---

Anti-pattern:

---
name: pdf-toolkit
# BAD - trigger info only in body (Claude can't see until AFTER trigger)
description: "PDF manipulation toolkit"
---

## When to Use  ← Claude never sees this before deciding to trigger!

Skill Anatomy

skill-name/
├── SKILL.md (required)      # Workflow + metadata (<5k words)
├── scripts/ (optional)      # Executable code (0 tokens when run)
├── references/ (optional)   # Docs loaded as needed
└── assets/ (optional)       # Output resources (never loaded)

Resource Types

Key insight: Scripts execute WITHOUT loading into context = 0 tokens.

Progressive Disclosure (3 Levels)

Result: 98% token savings vs loading everything upfront.

Skill Creation Workflow

1. Gather Examples

Ask user:

• What functionality should the skill support?
• Example usage scenarios?
• What triggers this skill?

2. Plan Resources

3. Initialize

scripts/init_skill.py <skill-name> --path <output-directory>

4. Write SKILL.md

Order:

1. Write description (with ALL trigger scenarios)
2. Add keywords
3. Write concise body (<5k words)
4. Reference bundled resources

Frontmatter checklist:

• name: lowercase, hyphens
• description: starts with "Use when...", includes ALL triggers
• keywords: discovery tags

Body checklist:

• Instructions (imperative form)
• Resource table (what to load when)
• Quick reference (common commands)

5. Validate & Package

scripts/quick_validate.py <path/to/skill>
scripts/package_skill.py <path/to/skill> [output-dir]

6. Iterate

Test on real tasks → identify gaps → update → test again.

Key Rules

References