Check Results Cheapest First

How should you verify AI-generated results efficiently? Check in order from cheapest: scripted checks, real-screen measurement, fact checking by search, then independent review by two different AI models just before release.

Check AI output starting with the cheapest method. Do not ask an AI about mistakes a machine can catch.

How to use it

Step 1, machine check: use scripts to find banned words, personal data and format rules. It costs nothing and catches the same mistakes every time.
Step 2, real-screen measurement: open the actual screen with browser automation and check phone width, horizontal scroll, console errors and button behavior.
Step 3, fact check: verify numbers and sources against search.
Step 4, independent AI review: only just before release, give it separately to two different models.
If an AI review gives generous scores, look at its evidence again first. There was a real case where it gave a high score based on content that was not in the work.
Set the completion criteria before you start, as sentences you can verify by eye. Example: pressing the copy button pastes the full prompt.

Prompts to copy

Evidence-first review

Review this deliverable. Evidence comes before scores.
1. For each problem, write "location / what is wrong / the evidence sentence quoted exactly".
2. Do not use content that is not actually in the deliverable as evidence. If you cannot find it, write "not confirmed".
3. At the end, give a verdict of only "Pass" or "Needs fixes".
Completion criteria: (put the 2 to 3 criteria set in advance here)