No Verification, No Green Light
The docs' hardest rule: if you can't verify an output yourself, don't let the agent proceed until IT has. Make self-checking part of done — run it, screenshot it, list what was checked. For anything client-facing, 'it looked right' has never once been a defense.
Why it works
The docs' hardest rule, and the one that separates toys from production: if you can't verify an output yourself, don't let the agent proceed until IT has. 'Generated' is a state; 'verified' is a standard. Making self-checking part of every task's definition of done is what lets a non-developer ship real work — the agent can read, run, and test what you can only eyeball.
How to do it
- Redefine done in your own head first: done = verified, not generated.
- Make verification part of the task, not an afterthought: '…then run it and show me it works.'
- Demand evidence in kind: a screenshot for pages, an output sample for scripts, a passing check for data.
- Ask for the checklist: 'list exactly what you checked' — unverified corners appear immediately.
- For client-facing work, verify one level harder: 'test on phone size, with empty data, with a long name.'
- Neither of you can verify it? Then it isn't done — descope, or find the person who can.
Copy this
prompt
Before you call this done: run it, verify the result yourself, and reply with (1) what you checked, (2) evidence — output or screenshot, (3) anything you could NOT verify. Unverified work isn't finished.
What to watch for
- 'It looked right' has never once been a defense — for you or the agent; evidence in kind or it didn't happen.
- Verification theater exists: 'I checked and it's fine' without artifacts is the claim restated. The checklist demand ('what exactly?') dissolves it.
Sources
Original: conon.aiDocumentation: code.claude.com