The Handoff Problem: Editing AI Drafts Without Starting Over
Lee Harris·
There is a version of AI-assisted writing where the model generates a draft and the writer rewrites it from scratch. This is common enough that some people have concluded it is normal. It is not a workflow. It is a way of using AI for inspiration and then doing all the writing yourself.
A proper handoff from generation to editing means the draft is structurally sound and argumentatively coherent before you open it for revision. The editing pass should be changing words and tightening sections, not rebuilding the architecture.

Why drafts require reconstruction
The most common cause is a brief that did not specify the argument. Without a stated argument, the model organizes the piece around coverage rather than position. It tells you everything there is to say about the topic in a reasonable order. That produces a draft that is technically complete and editorially dead. Restructuring it into a piece with a point is not editing. It is writing.
The second cause is generating a draft that is too long for the brief. Models fill available space. If you request 1,500 words and your brief only has substance for 800, the remaining 700 words will be elaboration, restatement, and transition. You are editing against material that should not be there.
The third cause is not specifying what sections the piece should have. The model will impose its own structure, which defaults to introduction, three points, conclusion. That structure is defensible and boring. If you want something different, specify it. A brief that says "open with the problem in the first paragraph, avoid a formal conclusion" produces output that matches. A brief that is silent on structure produces the default.
What the handoff looks like when it works
You open the draft and read it once for structure and argument, not for prose. Does it open with the right problem? Does the argument hold together from section to section? Does it end where it should, or does it end with a summary of what it just said?
If those questions have acceptable answers, the draft is worth editing. If they do not, stop. Go back to the brief, find the gap, fix it, and regenerate. The time spent editing a structurally broken draft is worse than the time spent on a better brief plus a regeneration.
Once the structure passes, read for voice. This is a different mode of reading. You are looking for hedging phrases, for symmetrical sentences that sound like they were generated by a model that wanted to be fair to all perspectives, for the places where specificity gave way to generality. Those are the editing targets.
I do the structure pass on screen and the voice pass out loud. Reading aloud surfaces rhythm problems that a visual scan misses. A sentence that looks fine often sounds wrong when spoken. That is the one worth fixing.
What you should not do
Do not combine the structure pass and the voice pass. They require different attention. When you check both at once, you find some of each and miss most of both.
Do not add new content during the editing pass. If the draft is missing a point you wanted to make, add it to the brief and regenerate the section. Grafting new content into an AI draft mid-editing creates seams, because the inserted passage will have a different register than the surrounding text. The result is a draft that sounds like two different writers.
Do not let editing become rewriting. If you are changing every sentence, the draft was not a draft. It was a starting point, which is a less useful thing. A draft you can edit in 30 minutes is worth more than a starting point that requires two hours of reconstruction. The difference is usually in the brief.
More from this topic
How to Handle Research When AI Gets Facts Wrong
AI is useful for structure and drafting, unreliable for factual claims. A practical approach to research-first workflows that use AI for synthesis without trusting it for sourcing.
What "Generic" Actually Means and How to Fix It Specifically
Generic is not the same as vague. It means wrong register, wrong specificity, wrong rhythm. A taxonomy of the most common problems in AI output and how to fix each one.
When AI Output Is Good Enough and When It Isn't
Not every piece of content requires the same quality threshold. An internal FAQ is different from a bylined article. How to make the call deliberately rather than by accident.
Why AI Writing Sounds Flat and What to Do About It
Not reassurance. Diagnosis. What AI models optimize for that produces generic output, and the specific editing moves that restore specificity and register.