Section 01
Inputs the agent can trust
A skill should say which inputs are required and which can be inferred from context. This prevents the agent from filling gaps with guesses.
- List required paths, URLs, IDs, or account context.
- State what to do if an input is missing.
- Keep optional context separate from required context.
Section 02
Tools and files to inspect
The best skills point the agent at the fastest truth path before it starts improvising.
- Name the first files or routes to read.
- Prefer repo-local helpers over generic shell exploration.
- Tell the agent when external lookup is useful.
Section 03
Completion evidence
Completion evidence is the difference between claiming the job is done and proving the job is done.
- Require the command, screenshot, route check, or diff that proves the result.
- Make the final answer name the evidence.
- Avoid proof that depends only on the agent's confidence.
Section 04
Stop rules
Stop rules keep the skill from expanding into unrelated work when it hits uncertainty.
- Name blockers that require a human decision.
- State what not to change.
- Tell the agent when to report partial progress.
Section 05
Feedback prompt
The feedback prompt should ask what happened, what rating fits the run, and what instruction would have helped.
- Ask for one rating.
- Ask for a short reason.
- Ask what to change in the skill next.