yjhqwer Verification Before Completion
Iron Law: NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE.
An unverified "done" is a guess wearing a confident tone. The proof is the check — run fresh and actually read — not the memory of having written the change.
The claim/proof table
| Claim | Requires | Not Sufficient |
|---|---|---|
| Tests pass | Test command run fresh, full output read, failures = 0 | "I wrote the tests" / an earlier green run |
| Build works | Build command exits 0 | "It looks right" |
| Bug fixed | The original failure scenario re-run, now showing the fix | Code inspection alone |
| Config/service works | The service probed after reload (health check, status endpoint, API call) | File edited + restart claimed |
| Subagent completed | Its report read AND its evidence independently spot-checked | Report says DONE |
| Requirements met | Each requirement mapped to its evidence | A summary paragraph |
Evidence scales to the claim. A one-line typo needs the changed line read back; a feature needs its test, build, or screenshot. What never scales to zero is some fresh, read output.
The gate
- IDENTIFY — What command or probe would prove this claim?
- RUN — Execute it fresh and complete, not "I ran it earlier".
- READ — Full output, exit code, failure count. Not the first line — the verdict.
- VERIFY — Does the output confirm the claim as stated?
- ONLY THEN — Make the claim, citing the evidence.
After proof passes, stop
Report commands, results, and unresolved risk — then stop. No polish, no cleanup, no unrelated tests after the criteria pass. Verification ends a task; it does not open a new one.
Extended from obra/superpowers verification-before-completion and juliusbrussee/caveman verify-and-stop (both MIT); wording adapted.