Editing a draft instead of rolling the dice again

The regenerate button is a dice roll

The draft came back not quite right, so you pressed "regenerate". You now have a second draft that is differently not quite right, and no information about why.

Understand what that button does. It runs the same prompt through the same model again and samples a different path. Nothing about your dissatisfaction was recorded. It cannot be, because your dissatisfaction was never expressed. Regenerating is a re-roll, and re-rolling is a reasonable thing to do twice when you suspect you got an unlucky draw, and a waste of time on the fifth attempt when the problem is your brief.

The alternative is to edit. Editing works — but only in a specific form, and the form is not the one most people use.

"Make it better" has no referent

Ask for "better", "more professional", "punchier", and the model has to guess what you mean. Its guess is drawn from what such requests usually mean in its training data, which is: longer, more adjectives, more structure, more hedging. You will get a version that looks like it tried harder. It will not be closer to what you wanted, and after three rounds of it you have a document made of connective tissue.

The instruction that works names three things: what to keep, what to change, and what the change is.

Keep paragraphs one and two exactly as they are. In paragraph three, delete the sentence beginning "We would suggest that" and replace it with a direct statement that the deadline is 14 March. Do not add anything else.

That is not a more polite version of "make it better". It is a different kind of instruction: it converts an aesthetic reaction into an operation on identified text. The model can execute it because every part of it points at something in the window.

Quote the thing you dislike

The single highest-value habit here costs five seconds. Paste the sentence you object to, and say what is wrong with it.

This line — "We are excited to explore how we might partner together going forward" — is doing no work. Replace it with one sentence stating what I want the reader to do.

Compare that with "the opening is too corporate". The first identifies a target; the second describes a mood. Models are literal readers, and a literal reader can act on a target.

Drift, and the fix for it

Edit the same draft seven times and something unpleasant happens. The good sentences from version one have been quietly rewritten along the way. Each turn re-generates the whole text, so nothing is protected unless you protect it — and "keep the rest unchanged" is obeyed loosely, because the model is producing the text afresh, not applying a patch.

Watch for the signature: each turn fixes one thing and breaks another, and you find yourself reinstating a phrase you liked two rounds ago.

The fix is to reset the anchor. Take the best version you have — assembled by hand from whichever paragraphs you actually want — paste it into a new message as the base text, and give one instruction against it:

Below is the current text. Change only the third paragraph, as follows: ... Return the full document with everything else byte-identical.

You have replaced a fourteen-message history full of superseded drafts with one clean input. It is cheaper, the attention is not spread across six wrong versions, and you know exactly what the model is working from.

Ask it what it changed

After an edit, a cheap audit:

List the changes you made, quoting the old text and the new text for each.

Two things fall out. You catch the silent edits — the number that became a rounder number, the qualifier that vanished from a sentence you had not asked about. And you get a record, which matters if the document is one somebody else will be accountable for.

Verify the quotes are real. If it claims to have changed a sentence that is not in either version, you have learned something useful about that particular reply.

Scolding does nothing

"That's wrong, try harder." "No, I already told you." "This is unacceptable."

None of these carry information. The model has no state that improves under pressure, and the previous rejection is only present as text in the window — text that now contains an argument, which will influence the next reply in ways you did not intend. Long conversations that have gone sour tend to stay sour, partly because the sourness is in the context.

If you are irritated, that is a signal that the brief was wrong, not that the model was disobedient. Start a new conversation with a better brief. It usually takes less time than the argument would have.

When to stop editing altogether

Three signals, and any one of them means stop:

• Your third edit has reintroduced the problem your first edit fixed.
• You have written more words of instruction than the document contains.
• You are editing a document you would have written differently from the first line — in which case the structure is wrong, and structure is not something an edit repairs.

The last one is the most common and the most expensive. A draft with the wrong shape is not a draft with problems; it is the wrong document, arriving quickly.

Seeing what actually changed, free

Comparing two drafts by eye is unreliable, and the changes that matter are usually the small ones. Every office suite will do it for you: LibreOffice Writer has Compare Document under the Tools menu, and Microsoft Word has Compare under Review. On a Mac or Linux machine the command line diff old.txt new.txt takes two seconds and shows you every altered line.

Run one comparison between a draft you accepted and the version you thought you had approved. Most people find at least one change they did not notice, and after that they keep the habit.