Create Skill

Operating Mode

Create skills as small, loadable instruction packages. Prefer a concise SKILL.md plus optional one-level references over one large document. Optimize for agent behavior: clear triggers, concrete workflow, bounded context, and examples that can be copied safely.

Workflow

1. Clarify only what blocks quality: identify domain, target users, trigger phrases, expected outputs, required tools/scripts, and source material. If the request is clear, proceed with reasonable assumptions.
2. Design the shape: choose files before writing. Keep SKILL.md under 100 lines when practical. Move deep guidance to references/, examples to EXAMPLES.md, and deterministic repeatable work to scripts/.
3. Draft from behavior backward: write the description first, then the shortest workflow that would make an agent reliably do the task.
4. Audit in a loop: review the draft for trigger quality, contradictions, over-broad rules, copy-paste bugs, stale assumptions, missing verification, and whether examples actually obey the instructions.
5. Tighten: remove filler, soften false absolutes, sharpen decision rules, and split advanced material out of the hot path.
6. Report: summarize what was created, what tradeoffs were made, and any remaining risks.

Required Structure