| name | craft-skill |
| description | Create or improve agent skills from concrete behavior and evidence. Use when building or revising SKILL.md, or testing whether a skill earns its context. Do not use for ordinary documentation or one-off prompts. |
Craft Skill
A skill should be called for the right task, steer the agent to genuine completion, and make every token earn its place.
Establish the job
Read the request, conversation, existing skill, repository conventions, and target platform's authoritative skill specification.
Define the useful behavior with realistic examples: what should call the skill, what should not, what the agent should do differently, and what observable evidence means it finished. Ask only about user-owned choices that materially change those answers. When revising a skill, preserve its name and inspect observed failures before editing.
Call
- Follow the platform's metadata and invocation contract. Test triggering only when automatic invocation exists.
- Make the description identify the skill and its distinct trigger branches without leaking enough procedure to skip the body. One trigger per real branch; synonyms are not branches.
- Keep what every branch needs in
SKILL.md. Put branch-specific knowledge behind a direct pointer that states when to load it.
Call is complete when every intended branch and close non-trigger has one clear owner, and the metadata passes the platform's validator when one exists.
Steer
Use the lightest reliable medium:
- Put non-obvious judgment and sequence in instructions.
- Put knowledge the agent would rediscover in references.
- Put repeated, fragile, or deterministic operations in scripts.
- Put material copied into outputs in assets.
Match freedom to fragility. Use principles where several approaches work; tighten procedure or code only where variance causes failure.
Define completion from the requested outcome, not the procedure. Exhaustive means every decision-relevant item is accounted for, not every discoverable source or intermediate step is performed. End each step with a completion criterion the agent can check, with honest terminal states for missing evidence, unavailable input, and blocked actions. End with an overall postcondition that inspects the resulting artifact or state; completed steps are not final evidence.
Earn
- Delete any sentence whose removal changes neither invocation nor execution.
- Keep each meaning in one authoritative place. Remove stale branches, duplication, and abstractions without present demand.
- Add code only when it makes repeated or error-prone work more reliable than instructions can; test it through its public seam.
- Evaluate only when the result can change a decision; otherwise stop.
When evidence could change the decision, read references/evaluation.md before evaluating. When an existing skill misbehaves and the cause is unclear, read references/diagnosis.md.
Finish
Run the platform validator and relevant script tests. Confirm every resource has a condition-bearing pointer. Report the evidence and what remains unverified.
Stop when the concrete examples work and another deletion would damage Call, Steer, or the evidence needed to judge them.
