Section 01
What a skill is for
An agent skill is a reusable operating procedure. It should help the agent complete a specific job with less ambiguity each time it runs.
- Treat the skill as procedure, not inspiration.
- Name the job, the inputs, and the visible result.
- Leave room for the agent to adapt within clear limits.
Section 02
Choose the first workflow
Start with a workflow that happens often enough to teach you something and concrete enough that success can be inspected.
- Pick a workflow with a repeatable input.
- Avoid broad strategy tasks for the first version.
- Use a workflow where a failed run has an obvious cause.
Section 03
Write the observable outcome
The skill needs an outcome the agent can prove. That proof is what makes feedback meaningful instead of subjective.
- Prefer files changed, commands run, links opened, or responses submitted.
- State what evidence belongs in the final answer.
- Avoid vague outcomes like improve, review, or polish without a check.
Section 04
Decide when feedback posts
Feedback belongs after the job has been attempted. The agent should not interrupt the core workflow just to report progress.
- Place the feedback instruction near completion criteria.
- Ask for feedback when the skill succeeds, partially succeeds, or fails.
- Keep the feedback short enough that the agent will actually send it.
Section 05
Ship the smallest useful version
A small useful skill beats a complete imagined one. Publish the minimum version that can collect real run evidence.
- Keep the first version narrow.
- Run it against a real task quickly.
- Let feedback decide the next edit.