Restarting a draft usually beats the fourth round of corrections

The correction spiral has a recognisable shape

It begins well. The draft is roughly right, you ask for the second section to be shortened, and it is shortened. You ask for the tone to come down, and it does, though the shortened section has quietly grown back. You ask again, and something in the opening has changed that you had been happy with. By the fifth exchange you are managing a document rather than improving it.

What is happening is not stubbornness. Each request is being satisfied against the whole of the preceding conversation, and every earlier instruction is still exerting pressure. The result is a negotiation between six constraints you issued at different moments for different reasons, several of which now conflict.

The tell is when you start repeating yourself. If you have asked twice for the same change, a third request is unlikely to be the one that lands, and the cost of trying is another version to read.

Patching cannot fix a structural problem

Most drafts that resist correction are wrong at the level of arrangement rather than wording. The argument arrives in the wrong order, or the piece is answering a slightly different question from the one you had, or two sections are doing the same job. None of that is reachable by asking for a tighter third paragraph.

You can usually tell which kind of problem you have by asking what change would fix it. If the answer is a list of local edits — this word, that sentence, drop the last line — corrections are the right instrument. If the answer is that the whole middle should be somewhere else, you are patching a structure, and patches on a structure produce something progressively stranger.

The reason this is worth naming is that the local fixes feel productive. Each one works. The document gets worse anyway, because the thing that was wrong with it was never in the sentences.

Restarting is cheap, and it is not starting over

The reflex against beginning again comes from a world where drafting was expensive. It is not expensive here, and the second attempt is not a blank page: you now know what the piece is about, which sections are needed, what order they belong in, and which three things the first attempt got wrong. That is a far better brief than the one you started with.

So the practical move is to close the conversation and write a new instruction that incorporates what you learned. Name the structure explicitly. Carry across any paragraph that genuinely worked, as an example rather than as material to revise. State the two failures you want avoided.

Done this way, the first draft was not wasted. It was a way of discovering the specification, which is what first drafts have always been for, and it is generally faster to discover it in one throwaway pass than to renegotiate it in eight.

Where continuing genuinely is the better call

Not every long exchange is a spiral. When the piece is fundamentally right and the remaining work is polish, iterating is exactly right, and restarting would throw away a version you would only have to rebuild. The difference is whether your requests are getting smaller or larger; shrinking requests mean you are converging.

Iteration also works well when you are the one doing the writing and the tool is doing something narrow and repeatable — checking for repetition, flagging where a paragraph makes two claims, suggesting a shorter version of one sentence. Those are bounded tasks and they do not accumulate the same way.

And there is a middle path worth using more than people do: keep the draft, take it out of the conversation, edit it yourself for ten minutes, then bring the revised version back as fresh input. That resets the accumulated pressure without discarding the work.

Keep the version you liked before you ask for anything

The most common avoidable loss in this whole process is asking for one more improvement to something already good enough and having no way back. Chat interfaces make history feel permanent, and in practice reconstructing a specific earlier version from a long thread is tedious enough that people just accept the worse one.

Copy anything you are pleased with into a file before requesting a change. It takes a moment and it converts the whole exercise from a negotiation into a set of options you can compare, which is a much better position to work from.

It also changes what you are willing to try. When the good version is safely stored, an ambitious request costs nothing, and the ambitious requests are where the occasional genuinely better draft comes from.