Asking an AI to review its own output sounds circular, and for facts it largely is. But there is a real and useful version of it. The trick is understanding the difference between checking something against criteria you supplied and confirming whether something is true in the world.
Good at checking against criteria you supply
Give it your requirements, then ask it to go through the draft against each one and say where the draft falls short. This works because the standard is right there in the conversation and the job is comparison rather than recall.
Poor at confirming facts
Asking whether the figures are correct produces reassurance generated the same way as the original answer. It has no external source to check against, so a yes tells you nothing. Verify facts against a real source, every time.
Ask for weaknesses, not approval
The question is this good invites agreement, because agreeable text is the likely continuation. Asking for the three weakest points, or what a sceptical customer would object to, produces genuinely more useful material.
A fresh look can beat a follow up
Paste the draft into a new conversation and ask for a critique without mentioning that it wrote the text. Without the earlier conversation pulling it towards consistency, the review is often sharper.
You are still the last check
Self review is a way of catching more of your own oversights before you look. It is not sign off, and it does not transfer responsibility. Anything leaving the business is your work and carries your name.
Common mistakes
- Asking are you sure and treating the reassurance as verification.
- Asking is this good, then taking the agreement as a quality check.
- Asking it to confirm figures or dates instead of checking a real source.
- Treating a clean self review as permission to send something unread.
- Reviewing against criteria you never actually stated in the prompt.
- Assuming a longer critique means a more rigorous one.
Check yourself
0/31.Which self review request is most likely to produce something useful?
2.You ask the model to confirm that the statistics in its answer are correct and it says yes. What have you learned?
3.Why can pasting a draft into a fresh conversation give a sharper critique?
This is general guidance, not a substitute for advice on your specific setup. Want a hand putting it into practice? Talk to us or see our care plans.