Why is asking again not a good idea?
Because you throw away what you already got. Typically not everything in the first answer is wrong: the structure is good, two of the three points are right, the tone is fine as well. If you start with a new question, you are gambling with those too.
There is another reason, and it follows from the mechanics. The model lays down tokens one at a time and fits everything to the text so far. If the first sentence set off in the wrong direction, the rest conformed to it — but you can see that in the answer. You can show it where things went off. A new question carries none of that information.
The three repair sentences
Almost every correction can be assembled from these three. You do not need all of them, but the order matters.
1. Name what is wrong — specifically.
Not "this isn't good", but: "the third point is wrong, because clause 4.2 of the contract says 30 days, not 15." A specific error tells it what to reconsider. General dissatisfaction only says try it differently — and the direction of "differently" will be random.
2. Say what to keep.
This is the most frequently omitted sentence, and it is the one that saves the good parts: "keep the structure and the tone, only rewrite the second paragraph." Without it the model tends to remanufacture the whole thing and break something that was fine before.
3. Hand over what was missing.
If the cause of the error was missing information, the shortest route is not an explanation but the material: "the relevant part of the framework agreement is here: [text]". That fixes in one step what three rounds of correction could not.
What not to ask
"Are you sure?"
This is the most common reflex, and it does not work. Within the same conversation the model is looking at its own earlier statement, and the most likely continuation is a confident confirmation. "I am completely sure" is worth exactly as much as the original claim — the same mechanism just ran once more.
If you really want to check, do it in a separate conversation, with the original material and without the previous answer. There it has nothing to defend. Why Does AI Make Things Up? covers this in more detail.
"Rewrite it, but better."
It carries no information. The model does not know what you consider better — if it knew, it would have written that the first time.
When is starting over worth it after all?
In three cases:
- When the first sentence set off in the wrong direction. Everything after it conformed to that. A mistaken basic assumption is harder to get corrected than to start over with a clean slate.
- When you have corrected three times and it has not improved. That almost always means the problem is not with the answer but with how the task was phrased. Go back to the prompt.
- When the conversation has become very long. Corrections accumulate, and the model has more and more botched versions in front of it. A clean conversation with the final material gives a cleaner result.
One correction, worked through
The first request: "Summarise these minutes for the management team."
What came back: seven points, two of them irrelevant, the decisions not highlighted, and one deadline stated wrongly.
A bad correction: "This isn't good, write it again properly, it's for management."
A good correction:
The structure and the length are good, keep those. Fix three things: 1. Points 4 and 6 are not relevant to management — take them out. 2. The deadline is not 15 March but 1 April — it is on page 3 of the minutes. 3. Pull the decisions out separately, at the top of the list, starting with "Decision:".
Those three sentences say exactly what should stay, what should go, and what was missing. Done in one round.
One habit that saves a lot
If the same error turns up a second time on a different task, then it is not something to correct but a missing rule. Put it in the prompt — or, if it applies to every task, in the system prompt. A correction is one-off; a rule solves it for good.
What to take away
Name the specific error, say what should stay, and hand over what was missing. Those three sentences settle most corrections. What not to do: don't ask whether it is sure — checking belongs in a different conversation.