How we test a skill
No badge without this page. A skill’s prompt is scored by AI reviewers against a checklist of what the finished result must contain, and the badge is read from that review, never typed in.
The question we ask
If you paste this prompt into ChatGPT, Claude, Gemini or Grok once, does it ask for everything needed to get the complete, finished result, with nothing left for a follow-up? We answer by having AI reviewers score the prompt against a fixed, published method. We do not rely on a lucky answer from one chat.
Layer 1: the result checklist
For each result, we write down what a finished deliverable from a skilled professional contains. For meeting notes that means every action with one owner and one date, the decisions, the open questions, the risks, and a follow-up email. Each item is required, or optional when only some people need it. A prompt covers an item only when it asks for it explicitly and says what good looks like. A passing mention does not count.
Layer 2: ten craft checks
The same ten checks for every result: the outcome and what “done” means, the inputs, the audience, the method, the exact output format, the rules, the edge cases, a self-check, whether it works in a single turn, and clarity. Each is scored 0, 1 or 2. They tell us whether the prompt will reliably produce what it covers.
Who scores it
AI reviewers score each prompt and write a line of evidence for every item. For most skills two reviewers scored blind, and a third decided any disagreement; five newer skills had one reviewer. Each review file names who scored it. These reviews read the prompt’s text. Reviewing real results is the next step, and each skill page says when that is still pending.
The ship rule
A prompt ships only when every required checklist item is covered, the three key craft checks (outcome, output format and self-check) each score 2, and the craft total is at least 16 out of 20. Miss one required item and it does not ship, however well it is written. The badge then reads “Reviewed · covers 10 of 10” with the date, and the checklist sits beneath it so you can see exactly what you will get.
The beat rule
Our aim is for each skill to score at least as high as the best public prompt we found for the same result, on coverage first and craft second. We adopt patterns from the best prompts, never their wording, and we credit where ideas came from.
The owner’s spot checks
Our policy is that the owner reviews every Care skill (the ones that carry a safety block) and a random sample of the rest before it goes live. A skill says “Spot-checked by the owner” only once that check is recorded in its review, with the date. Reviews are saved next to each skill with the scores, the evidence, the date and the version of the method, so no badge is ever typed in by hand.
Not reviewed yet
A skill with no passing review says “Not reviewed yet”, stays out of search engines, and is listed after reviewed skills in the library. You can still use it. The checklist under “What you’ll get” helps you check the result yourself.
Never invented
We show a sample result only when we have a real one. Until then you see the checklist of what the prompt covers, not a made-up answer.