Same brief, same model. Published either way.

The brief is written from the skill's own SKILL.md, so it tests what the skill says it does, and then it is fixed. Authors can suggest a task, but they cannot write or edit the brief. It is shown on every proof page so anyone can repeat the run.

Claude builds the brief three times without the skill and three times with it, on the same model and settings. All six results are published, so you see the spread and not one lucky draw. Outputs are never edited.

Every proof records the skill's repository, the exact commit it ran against and the date. If the skill changes later, the proof still describes the version it tested.

Authors can pay to have a skill proven. Those proofs are marked Author-proven. A paid proof is published whatever it shows, under the same rules as every other proof, and an unflattering result is not refunded. We refund only when a proof cannot be completed for a technical reason.

The README badge links to the proof. Twelve months after the proof date it changes to Re-run due, until the skill is proven again on a current commit.