Creating Agent Skills

Create skills for reusable, task-specific workflows that need procedural guidance, domain knowledge, tools, or packaged resources.

Use AGENTS.md for durable repository conventions and prompt or thread context for one-off instructions. Do not create a skill merely to restate behavior the host already provides reliably.

Treat every skill as a behavioral artifact: define its activation boundary, intended behavior, evidence of success, and safe failure behavior.

Workflow

1. Resolve the target hosts, installation scope, expected user requests, non-trigger near misses, dependencies, and safety boundaries.
2. For an existing skill, read its complete SKILL.md and directly referenced resources before proposing changes.
3. Identify the smallest reusable workflow and decide whether instructions alone are sufficient.
4. Add references/ for on-demand knowledge, scripts/ for repeated deterministic work, and assets/ only for files used in outputs.
5. Write a concise description that distinguishes this skill from adjacent skills.
6. Define expected inputs, workflow steps, outputs, unsupported inferences, and ask, stop, or fallback conditions.
7. Validate the skill against the target host or specification.
8. Test activation and task behavior with representative positive, negative, incomplete, and edge cases.