How to Review AI Drafts

Review against a checklist that existed before the chat

Most AI drafts are reviewed by rereading them for tone. That is how a fluent page with the wrong ask, the wrong reader, and an unsigned risk still gets sent. The review has to start before generation. Write the audience, the ask, the facts in play, the risk if a sentence is wrong, and the one sentence you are willing to sign. Generate only after that list exists. Then read the draft against the list, not against itself. If the model wrote a smoother version of a decision you have not taken, that is not a good edit. That is a new decision wearing your grammar. You still own the decision. The checklist is how you notice the swap.

A board letter and a Slack update can share facts and still be different artefacts. ChatGPT will happily turn either into a generic professional note. Your review is whether the reader is named, whether the ask is in the first short block, and whether every number still matches the pack. Voice is the last pass, not the first. Put back the nouns the model sanded off. Cut the throat-clearing that arrived to sound complete. Restore the judgement you actually hold, including wait and no. If you would not say the sentence in the meeting, it does not survive the review, however balanced it looks on the page.

This essay is a method for reviewing drafts you will send under your name: emails, memos, commentaries, and notes to a client. It is not a style guide. Style without a checklist is how you polish a wrong ask. A named human still signs. If nobody can say who is accountable when a line is wrong, you do not have a review process. You have a faster way to produce unread pages. Write the checklist. Then generate. Then edit in the document that will actually leave the building every time.

Audience, ask, facts, risk, and the signable line

Five items are enough. Audience: who will read this and what they already know. Ask: the one thing they must do or understand. Facts: the pack you will allow, with bans on new numbers. Risk: money, legal position, reputation, or a named person. Signable line: the sentence you would read aloud with the file open. Generate against that list. Review against that list. If the draft invents a second ask, cut it. If it softens a no into a maybe, restore the no. If it adds a courtesy promise, delete the promise. The model is trying to be helpful. Helpful is how extra commitments appear. Your review is a reduction to what you already decided.

Edit in the artefact that will be sent, not only in the chat window. Copy-paste from a blessed chat into Outlook is how an earlier error returns. Check names against the thread. Check dates against the calendar. Check figures against the cell. Then do a voice pass: two paragraphs you actually wrote, compared with the draft, to put back your cadence. Do not ask the model to write in a punchy world-class voice. That request produces brochure English. Ask it to keep your nouns and shorten. You still cut. Review is labour. The time you recovered in generation should be spent here, not on another generation.

Review once against the list, then send or stop

Keep the checklist in the same note as the brief so you cannot generate without it. After one generation, score the draft: audience right or wrong, ask present or missing, facts sourced or invented, risk handled or expanded, signable line present or absent. Fix the fails yourself where they are judgement. Send the model only mechanical fixes: shorter, same nouns, mark [missing] on unsourced lines. Then you read the sent artefact start to finish. If a second generation is needed, change one instruction, not five. A loop of regenerate until it feels done is how the checklist dies and the vibe returns.

For recurring mail, save a reviewed brief rather than a reviewed personality. The weekly status note, the polite chase, the decline with one alternative: those can be templates with the five items filled. A new person on the team should be able to run the review without asking you what good looks like. If the only standard is I will know it when I see it, you will rubber-stamp fluent pages when you are tired. Tired is when unsigned risk ships. The checklist exists for Tuesday evening, not for the demonstration on Monday morning.

Role: You are an editor reviewing a draft I will send under my name.
Task: Score the draft against my checklist and propose only mechanical fixes.
Context: I will paste the checklist I wrote before generation, plus the draft.
Constraints:
- Do not add facts, promises, dates, or a second ask.
- Do not change my judgement from no or wait into a softer yes.
- Mark [missing] on any line that is not supported by the checklist facts.
Output: Pass or fail on audience, ask, facts, risk, and the signable line, then a tightened draft.
Quality checks: What commitment a reader could think I made that I did not list.

Move the surviving draft into the email, memo, or slide that will actually be sent. Check names, dates, and figures there. Read the signable sentence aloud with the file open. If you would not say it, change it yourself. Do not ask the model for a braver version of a decision you have not taken. Save the checklist that caught a miss. That list is the review asset, not the chat transcript.

Tone-only review, extra promises, and rubber stamps

Tone-only review rewards the model's strongest skill: sounding finished. Extra promises hide in courtesy. Happy to send a revised timeline is a commitment if you send it. Rubber stamps happen when a manager glances at a fluent page and forwards it. If you manage people who use AI, require the checklist in the pack, not a screenshot of the chat. Ask them which sentence they will sign. If they cannot answer, the draft comes back. Volume of drafts is not a performance metric. Unowned sentences are a risk metric. Your name on the forward is the name that will be asked when the figure is wrong.

Watch voice theft that still fails the facts. A draft that sounds like you and cites the wrong period is worse than an awkward draft with the right period. Colleagues will trust the voice. Put the voice last. Also watch confidential review. Pasting a labour letter into consumer ChatGPT so the model can help you review it is a classification error. Review in the approved tool, or review with your eyes and the file. The checklist still applies. The tool does not replace the five items. It only changes where the draft sat while you ignored them.

Related reading on StudyGrid

Read next: Writing Someone Will Sign ChatGPT for Email at Work How to Write Better ChatGPT Prompts. Those essays sit beside this one. Use them when you need the neighbouring skill, not as a substitute for the check you still have to make.

What to do this week

Before you generate anything you will send, write the five items on a note. Generate once. Score the draft against that note. Fix judgement yourself. Move the text into the artefact and check names, dates, and figures there. Send only if the signable sentence still holds with the file open. Keep the checklist that caught a miss so the next review is faster, not looser. Do this on one real send this week.