Editing a generated draft is not proofreading. Proofreading fixes what is wrong; editing generated text fixes what is missing. The sentences are usually correct, the paragraphs are usually coherent, and the piece still does not say anything a reader could not have guessed.
The work is different enough that the order matters. Most people edit line by line from the top, which is the slowest possible approach and the one that leaves the structural problems in place until the end, when fixing them means undoing the line edits.
Three passes, done in a fixed order, take about the same time and produce a materially better piece.
Editing is not proofreading
A generated draft arrives with a specific profile. Its structure is conventional to the point of being predictable, its claims are hedged uniformly, and its examples are plausible without being checkable.
Line-level editing does not touch any of those. You can polish every sentence and still have a piece whose second half adds nothing to its first, because the problem was never at the sentence level.
So the first question is not "is this sentence good" but "should this section exist". That question has to be answered before any sentence-level work, or the sentence-level work is wasted.
Pass one: structure
Read only the headings. Do not read the paragraphs underneath them yet.
Ask, for each heading: could a reader predict what this section says without reading it? If yes, the section is likely doing nothing. "Benefits", "Challenges", "Best Practices", "Conclusion" are the usual offenders, and their content is almost never worth the space it occupies.
Then check the order. Does each section depend on the previous one being understood? If two sections could be swapped with no change in meaning, they are parallel rather than sequential, and the piece has no argument — it has a list.
Finally, look at the ending. If it restates the introduction, cut it and write a real one. This is the single most common structural defect and the cheapest to fix at this stage, when nothing has been invested in the prose yet.
Pass two: specificity
Now read the body, looking at one thing only: whether each claim is anchored to something checkable.
Mark every sentence that asserts something about the world. Then mark how many of those sentences contain a number, a named example, an observed detail, or a stated condition. Generated drafts typically have very few, and the few they have are often invented.
For each unanchored claim, you have three options. Add the specific detail you actually know. Add the condition under which the claim holds, which converts a general assertion into a usable one. Or delete the sentence.
Deleting is underused. A piece that loses a fifth of its sentences and gains a clear argument is better than the same piece kept at full length.
Pass three: voice
Voice is the last pass, not the first, because changing the wording of a section that should not exist is a waste of the effort.
At this stage you are looking for the tics that mark generated text: the same sentence length repeated, transitions that say nothing ("moreover", "furthermore", "it is worth noting"), hedges stacked in the same pattern, and paragraphs that all open with a clause rather than a subject.
Fix them by varying rhythm rather than by substituting synonyms. The most effective single change is to take one long sentence in each paragraph and split it into two, or to merge two short ones. Rhythm is what readers perceive as authorship, far more than word choice.
If you have a style guide, this pass is where it applies. If you do not, this pass is where you discover you need one, because you are making the same corrections for the fifth time.
Why the order matters
Doing voice first means polishing sentences in sections that will later be cut. Doing specificity first means hunting for concrete details in a section that may not survive the structure pass.
Structure first also changes what the other passes cost. Cutting two sections reduces the body you have to work through by a third, and it often removes the worst instances of hedging and the vaguest claims, because those concentrate in the filler sections.
When time is short, run the structure pass only. A piece with a sound structure and unpolished prose reads far better than the reverse, and it is closer to publishable than it looks.
What not to fix by hand
Some corrections are worth making once at source rather than repeatedly in the draft.
If you delete the same phrase in every piece, add it to the banned list in your prompt or style guide. If you always have to add the same missing element — a limitation, a next step, a source for a claim category — add it to the template. If you consistently rewrite the opening, the template's opening instruction is wrong.
The test is repetition. A one-off correction takes thirty seconds and should just be made. A correction you have made five times is a process defect, and fixing it once at the source removes it from every future piece.
Deciding when a draft is done
Generated drafts never feel finished, because there is always another sentence that could be improved. Without a definition of done, editing expands to fill whatever time is available and the marginal return collapses.
A usable definition: the structure passes the heading test, every claim has an anchor or has been deleted, the rhythm varies, and there is nothing in the piece you would be uncomfortable defending if a reader asked about it.
That last condition is the real one. Editing is finished when you would sign the piece, not when it is perfect — and the difference between the two is smaller than most people assume, while the time difference is usually a factor of three.
Building a house style from your edits
Keep a running list of what you change, in two columns: the thing you removed, and the thing you replaced it with. After twenty pieces the list is a description of your publication's voice that no style guide could have produced from scratch.
That list is also the most reliable input for improving prompts. A template enriched with a list of banned constructions and required elements outperforms one enriched with adjectives, because a ban on "in today's fast-paced world" is a specific, checkable instruction, while "write engagingly" is not.
Over time the goal is for the standard to live in the process rather than in the editor's attention. Every correction that moves from your head into the template is a correction you no longer have to make, which is what makes it possible to publish more without lowering the bar.