Paste a draft into an AI tool and ask, “Is this good?” and you have left the hardest part of the review undefined. Good for whom? Judged against what? Looking for which kind of problem?
A useful review needs a reader, a failure mode and a standard. It also needs a human who is willing to challenge the answer.
Why broad review prompts underperform
“Is this good?” leaves the tool to invent the audience, the standard and the depth of review. The result can sound reassuring while telling you almost nothing.
Give it a narrower job. Ask for the weakest claim, the unanswered objection, the line a hurried reader could misread or any number without a source. Then ask it to show its working by quoting the relevant line.
How to run an adversarial review
- Protect the material. Before sharing client or confidential work, check the provider, account tier and settings. Find out whether the input is retained or used for training. If the terms are unclear, remove sensitive details or use a non-confidential example.
- Name the reader. Give the tool a sceptical role and a reason that person might reject the work.
- Name the failure. Ask it to look for unsupported claims, missing boundaries, unanswered objections or ambiguous instructions.
- Demand evidence. Put a direct quote from the draft beside every criticism.
- Review the review. The tool can surface a question. It cannot decide whether that question matters in your situation.
The proposal stress-test uses the same method on one document.
Where this fits in your process
This is the checkpoint between “I think it is done” and “it is sent”. If the underlying process is unclear, the review may describe the mess without repairing it.
Common questions
Does this make the review negative for the sake of it?
No. The role gives the review a point of view. You still decide what is useful and what is noise.
Can I use the same instruction for everything?
Keep the structure, but change the reader, the failure modes and the standard. A proposal, a client email and a project plan do not carry the same risks.
Which tool should I use?
Use a tool that meets your data-handling requirements and test it on your own material. The instruction and the human review matter more than the logo on the screen.
Before you press send
Ask for a review from the reader's side, then verify every criticism yourself. If you want to map where checks like this belong across the wider process, book a Clarity Session and work through it directly with Shaun.
