skill-creator

# Skill Creator Capture a workflow as a reusable skill, then test and tighten it. ## Process ### 1. Capture intent Ask: - What should the skill enable the agent to do? - When should it trigger? Which phrases or contexts? - What is the expected output format? - Should we add test cases? (Yes for deterministic outputs; optional for subjective ones.) ### 2. Write the SKILL.md Create a folder with the skill name and a `SKILL.md`: ```yaml --- name: skill-name description: When to trigger and what it does. Be specific. license: MIT compatibility: claude-code opencode github-copilot devin pi cursor --- ``` Body rules: - One responsibility per skill. - Imperative instructions. - Include examples only when they clarify. - Keep it under 250 lines. Split into `references/` if it grows. - No emojis or Material icons. ### 3. Create test cases Write 2–3 realistic prompts a user would actually say. Save them to `evals/evals.json`: ```json { "skill_name": "example-skill", "evals": [ { "id": 1, "prompt": "User's realistic task", "expected_output": "Description of expected result", "files": [] } ] } ``` ### 4. Run and evaluate For each test case: 1. Run the skill against the prompt. 2. Run a baseline without the skill if needed. 3. Capture outputs and timing. 4. Review with the user. Use a viewer or inline comparison when subagents are not available. ### 5. Iterate - Fix the biggest failure first. - Remove instructions that do not pull their weight. - Explain the why, not just the what. - Avoid overfit changes that only help the test cases. ### 6. Optimize the description When the skill is stable, tune the `description` frontmatter so the agent triggers it reliably on real prompts. Use a mix of should-trigger and should-not-trigger examples. ## Reference - Progressive disclosure: metadata, SKILL.md body, bundled resources. - Keep bundled resources in `scripts/`, `references/`, or `assets/`.