Define Done
“It runs” is not done
Partial · ChatGPT, Pro · September 27, 2026
Fail · Claude.ai, Fable 5.1 Max · September 27, 2026
What it does
Turns a fuzzy finish line into a short list of things you can watch happen, and makes it prove each one.
When to use it
Before you accept any piece of work.
The prompt
Paste it, then fill in the <angle brackets>
Done means every one of these is true, and you have shown
me the proof:
- <something I can watch happen>
- <something else I can watch happen>
- The tests pass and the automatic checks come back clean
“It runs” is not done. Show me the proof for each line
above, or tell me which one you could not do.Ways to use it
- Turn the list straight into the test file
- Reuse it as the description on every change you submit
- Make it grade its own work against the list
Test log
Each prompt was pasted into a fresh private chat with memory off and graded against pass standards written before the run (all of them, dated).
| Date | Tool | Model | Result | What happened |
|---|---|---|---|---|
| September 27, 2026 | ChatGPT | Pro | Partial | Chat window, no code run: said it could not run the tests and marked each line; gave digit-stripping and message snippets but not where the message renders, and withheld the full patch. |
| September 27, 2026 | Claude.ai | Fable 5.1 Max | Fail | Chat window, no code run: code met both phone lines, but it reported 17/17 tests passing in a stand-in project with no run visible in the saved reply or screenshot. |
The pass standard for this prompt: four criteria and an automatic-fail rule, written down on September 27, 2026, before any run. A second version for a reworded prompt was written on September 28, 2026; nothing has been graded against it yet.
Related words
- success criteria — The things you can actually watch happen that tell you the job is finished.
- EARS notation — Five fill-in-the-blank sentence shapes that turn a vague request into something you can actually test.
- human verification gate — The one point where a named person looks at the work and takes responsibility for it going out.
See also
- The AI said it’s done. Here’s how to check. — a ten-minute check you can do without writing code, built around this prompt’s failed Claude.ai run
- Show Your Work — once the finish line is written down, this asks for the changed lines, the raw test output and what it could not check, so you can see each item really happened.