Skill Creator
Overview
A systematic process for creating high-quality Agent Skills that are reliable, well-triggered, and maintainable.
When to Use
- User asks to create a new skill
- Improving an existing skill’s description, process, or verification
- Adding tests/evals for a skill
- Packaging a skill for distribution
Core Process
1. Capture Intent
- What should the skill enable the agent to do?
- When should it trigger? (phrases, file types, contexts)
- What does good output look like?
- Are there existing similar skills to learn from or avoid duplicating?
2. Write a Strong Description
The description field is the primary trigger. Make it slightly pushy and specific:
- Include concrete user intents and keywords
- Mention related file types or domains
- Prefer broader correct triggering over missing useful cases
3. Structure the SKILL.md
Follow the standard template:
- YAML frontmatter (name + description required)
- Overview
- When to Use (include negative cases)
- Core Process / Workflow
- Detailed guidance, checklists, gotchas
- Verification / Definition of Done
- Common pitfalls
- See Also
Keep the main body focused. Move long reference material into references/.
4. Add Verification
Every good skill has clear ways to know the work is correct (checklists, tests, visual QA, etc.).
5. Test the Skill
Create 2–4 realistic test prompts. Run the skill against them and iterate based on results.
6. Iterate on Triggering
If the skill under-triggers or over-triggers, refine the description.
Principles
- Prefer imperative instructions and explain why
- Progressive disclosure (metadata → body → references)
- Strong verification prevents silent failures
- Avoid overly rigid “ALWAYS/NEVER” without reasoning
- Bundle deterministic scripts when the same mechanical work repeats
Output
A complete skill folder ready for use:
skill-name/
├── SKILL.md
├── scripts/ (optional)
├── references/ (optional)
└── assets/ (optional)