How to Tell Whether an AI Answer Is Any Good

“Looks good” is not a test. I need criteria that match the job. A code answer should run and satisfy the contract. A research answer should trace important claims to sources. A game concept should be playable under the stated constraints. For factual answers, Research Prompts That Separate Evidence From Inference adds a check that separates sources from the model’s own inference.

For a prompt asking AI to compare two storage choices for a game save system, I might use four checks: Correct behavior: Does it explain what survives a refresh and what survives a device change? Tradeoffs: Does it compare implementation effort, failure modes, and user expectations? Assumptions: Does it state whether players have accounts? Next action: Can I decide on a reversible first implementation? I can mark each check pass, fail, or unclear. “Unclear” matters; it tells me what to ask next rather than letting a confident paragraph stand in for evidence.

If I am picking a harmless theme for a prototype, originality and playability may matter most. If I am reviewing code that handles player data, correctness and risk matter more. Avoid applying a universal score to unrelated tasks.

For code, run tests. For factual claims, open the cited source and see whether it really supports the statement. For a public article, check dates and links. For a UX idea, test it with someone who has not heard the explanation. OpenAI’s evaluation guidance describes evaluation as a way to judge model behavior against examples and criteria; the same principle works in a small manual notebook.

If the answer lacks a next action, ask for one. If it cites a source that does not support the claim, ask for a corrected source-backed answer. Do not merely request “more detail”; target the failed criterion. Try it: Before your next AI request, write three pass/fail checks. After the result, mark each one. If one fails, ask for a revision that fixes only that check and rerun it. Keep learning: Research Prompts That Separate Evidence From Inference, Spot a Confident Wrong Answer Before You Share It, and How to Test a Prompt on More Than One Example.