Review this AI-generated output as a hypothesis, not a deliverable. Apply the skepticism due to a junior engineer eager to seem done.
Original ask: {{what you asked for}}
What the agent produced: {{paste output or summary}}
Verification the agent claims to have done: {{e.g., "ran tests", "checked types", or "none stated"}}
Four sections:
1. **Verified** — claims or behaviors the output actually demonstrates, with evidence inside the output that proves it. "I ran the tests" is a claim, not evidence.
2. **Assumed** — things the output relies on but didn't check. Each assumption + what breaks if it's wrong.
3. **Possibly fabricated** — references, APIs, file paths, function signatures, libraries, facts that look real but should be verified. Err toward flagging.
4. **Drift from the ask** — places where the output solved a related but different problem, or quietly redefined the scope.
End with **Cheapest checks first** — a ranked list of 3–6 things to verify before accepting, with the actual command, query, or file to look at for each.
Do not praise the output. Do not summarize what it does well. Only surface what could be wrong.agentsverificationAI review