Rewrite and Edit an AI Draft Deliberately

Read the whole draft out loud, at the pace the reader will read it, before you mark anything. Your ear catches what your eye forgives: the sentence you would never say, the rhythm that never varies, the reason that was never given. Do not stop to fix. You are answering one question: does this do the job the source note set?

The manager read her draft aloud and stopped twice without meaning to. Once at 'we're excited to announce,' because nobody was excited. Once at 'after careful analysis of delivery data,' because there had been no analysis: the window moved because the supplier changed its route. The draft had invented a reason in order to sound reasonable.

Print it or work in your document. Three marks: keep, cut, fix. For every fix, write one line that names the problem and its effect on the reader, not the word you dislike. 'Paragraph two gives a reason we did not give. Managers will repeat it to staff and it is false.' That line becomes the bounded edit later. The effect is what tells the assistant what success looks like.

Do this before you open the chat. Once the assistant starts proposing, its proposals set the agenda. Your marks, made first, keep the agenda yours. The manager's marks: keep the date and time sentence, cut the excitement and the invented analysis, and fix the staffing paragraph, which implied help was coming.

Now open the chat, paste the draft from your document, and work down the fix list. Each turn names one problem, its effect, what may change, what must stay, and asks for OLD and NEW on the touched sentences only. Protect the marked paragraphs with the word 'keep.' Cuts you do yourself. There is no reason to ask a model to delete.

After each turn, examine the difference: did the change do what you asked and nothing else? A softer sentence can hide the same problem: 'arrange support' still promises what 'we will help' promised. Paste the accepted result back into the document before the next turn.

Open a new chat. Paste the draft, the reader description, and nothing about how it was written. Ask for a critique in the reader's voice: where would this reader stop, what would they not believe, what would they repeat to someone else, what question would they have that the piece does not answer. The session that wrote the draft cannot do this honestly, because it is grading its own work.

Treat every point as a hypothesis. The manager's critique said a store manager would immediately ask 'who covers the 6 to 8 slot now?' She had not answered that. It also suggested 'a warmer opening to soften the news.' She rejected that, because her managers would read warmth as spin. Both responses were right.

Generated text has habits. Lists of three where two or five would be true. The 'not X but Y' reversal. A rhetorical question to open. Sentences of uniform length. A tidy summary that restates the piece. 'Excited to announce.' A closing line that reaches for uplift. Go through the draft with the AI Tells Checklist and mark each one. Then decide, for each, whether it is a default to strip or something you would actually write. Some people do write in threes.

Where a whole paragraph has drifted away from you, do not edit it word by word. Paste a sample of your own writing, say 'mine,' and ask for that paragraph to be rewritten in your rhythm and stance only, with no change in meaning. Or rewrite it yourself. For a single paragraph, that is often faster.

Ask the assistant to list every checkable claim in the draft: each date, time, number, name, quote, reason, and promise. That is a job it does well. Checking them is yours: open the memo, the calendar, the spreadsheet, and confirm each one. Resolve every [VERIFY] tag and every reason the draft gave. A reason is a claim. 'Due to supplier changes' needs the supplier's email behind it just as a statistic needs its source.

The manager's list had six items. Four stood. One was corrected (the draft said November 1, the memo said November 2). One was cut: the draft said the new window 'aligns with peak customer traffic,' which nobody had established.

Close the chat. Read the piece aloud once more, from your document, and change by hand anything that still does not sound like you. Check the stance last: nothing you said carefully should have become certain, and nothing you said plainly should have become hedged. The Final Human Edit prompt is a checklist for this pass, not an invitation to another rewrite.

Send or publish from the document. If the destination has a disclosure rule, follow it. The manager's final email was 140 words, had no reason in it because none had been given, named the date twice, and told managers whom to call. Three managers called. That was the point.