No "It Should Be Fine Now"
No "It Should Be Fine Now"
superpowers core skill · verification-before-completion
In one line: Don't accept the AI saying "done" without verifying. Make it run the command, read the output, and conclude with evidence.
Scenario: the AI says "the login timeout bug is fixed"
You ask the AI to fix "login occasionally times out at peak." Soon it replies: "I optimized the DB query, should be fixed." Note that "should." Trust it and ship, and peak traffic times out again — users complain, you take the blame. It says "fixed," but did it verify?
How to ask the AI
Don't just say "should be fine." Give me the evidence:
run the relevant tests and paste "X/X passed"; for this timeout, how did you reproduce it, and does it still time out now — paste a before/after.
| ❌ Don't accept | ✅ Demand this |
|---|---|
| Optimized the query, should be fixed | Load-tested 200 concurrent: timeouts before, 0 after — here's the data |
| I'm confident it's fine now | Ran the tests, 34/34 pass, output pasted |
| It's done | Checked the git diff, the change is really there, tests pass too |
Watch the two laziest spots: "optimized" needs before/after numbers; "tests pass" isn't proof this timeout is gone — make it verify against the original symptom.
How it'll walk you through it
It runs five steps: figure out which command/repro proves "no more timeout" -> actually run it (not by feel) -> read the full output and timing data -> does the data really support "fixed" -> only then say "fixed," with a before/after. Any "should / looks like" without evidence should, by the rule, stop and verify first.
Remember one line
verification-before-completion = make the AI talk with evidence. When it says "fixed / should be fine," ask "did you reproduce it? where's the data?" — never accept a completion claim with no proof.