Skip to main content
A skill is ready for use when its package is structurally valid and its behavior has been tested against representative requests. Publishing is a separate decision.

​
Manual lifecycle

1

Create the draft

2

Write representative tests

Add evals/routing.json with requests that should and should not activate the skill. Add evals/evals.json with two or three realistic end-to-end tasks, expected outputs, and concrete assertions.
3

Run the output comparison

Inspect both arms and the generated benchmark.json. A structural pass is necessary, but it does not prove that the instructions improve the work.
4

Verify the package

Review any generated evidence before accepting it. Use --no-auto-fix when you want verification to report missing artifacts without creating them.
5

Review before distribution

Confirm the package contents, permissions, external tools, failure modes, and validation report. Publishing or global installation requires a separate approval.
Agents should follow Improve and validate as an agent and keep installation or publication as separate approvals.

​
Success criteria

  • The description states both capability and activation conditions.
  • Instructions define the result, important constraints, and failure behavior.
  • Positive, negative, and realistic tasks are represented.
  • Output-quality cases run against a no-skill or previous-skill baseline.
  • Verification identifies the exact skill version it tested.
  • The package remains local until you approve publication or installation.