Skill Creator

Skill Creator

Workflow

1. Collect three concrete user requests that should trigger the Skill and two nearby requests that should not.
2. Decide which knowledge belongs in concise instructions, which belongs in on-demand references, and which deterministic work deserves a script.
3. Read references/package-format.md and create the canonical directory layout.
4. Write SKILL.md frontmatter with only name and description. Put every trigger condition in the description because the body is loaded only after selection.
5. Write imperative instructions for another capable agent. Include only non-obvious procedure and domain constraints.
6. Put UI metadata in agents/openai.yaml. Its default_prompt must mention $skill-name explicitly.
7. Keep references one level from SKILL.md and explain when each reference should be read.
8. Add scripts only for deterministic or repeatedly rewritten operations. Scripts must have bounded inputs, clear errors, and no embedded secrets.
9. Run scripts/validate_skill.py <skill-directory> in an environment that permits local scripts. On MiaoMiao cloud V1, packaged scripts are intentionally not executable; use repository CI instead.
10. Test the Skill on realistic requests without leaking the intended answer into the test prompt.

Quality Bar

• Prefer a small, precise Skill over a general prompt collection.
• Do not duplicate detailed reference content in SKILL.md.
• Never require direct execution of third-party scripts as a condition for basic use.
• Never place tokens, credentials, private keys, or environment files in the package.
• Use lowercase letters, digits, and hyphens for the directory and Skill name.
• Keep SKILL.md below 500 lines and split details into references before it becomes difficult to scan.
• Remove template placeholders, unused directories, generated caches, and auxiliary process documentation.