Name what is wrong before changing anything, then change one variable so you learn something from the result.
Do now: Take a disappointing output and write one sentence naming the specific gap.
'Make it better' teaches you nothing. 'Make it shorter and drop the second example' teaches you something.Iteration is the skill that separates people who find these tools useful from people who find them disappointing. It is also the cheapest skill to learn, because it costs nothing but attention.
Diagnose before you rewrite
When an output misses, the instinct is to rewrite the whole request. Resist it, and instead name the specific failure: too long, wrong audience, invented a fact, ignored a constraint, right content in the wrong format, or correct but lifeless.
Each of those has a different fix. Wrong audience is a briefing problem — say who is reading. Invented a fact is a supply problem — provide the fact or ask for a marked gap. Ignored a constraint often means the constraint was buried in the middle of a long paragraph rather than stated plainly at the end.
'Correct but lifeless' is the most common and the least diagnosed. It usually means you asked for content and forgot to ask for a voice, and the fix is samples of your own writing rather than more adjectives.
Name the gap in one sentence. If you cannot, you do not yet know what to change.
Change one variable
Change the format, keep everything else. Change the audience, keep everything else. Changing five things at once may produce a better answer, but you will not know which change did it, and next week you will be guessing again.
Keep what worked explicitly. 'Keep the structure and the first two points, replace the third with something drawn from the document, and cut the closing paragraph' preserves your progress. Starting over discards it, which is why long conversations often drift away from a good version that existed twenty messages earlier.
Save the version that worked before trying to improve it. This costs four seconds and prevents the specific misery of having had it right and iterated past it.
Know when to stop
Decide in advance what good enough looks like — the length, the format, the facts it must contain, and the things it must not claim. Without that, iteration continues until you are tired rather than until the work is done.
Two rounds of no real improvement is a signal. Either the request has a problem no rewording will fix, or the remaining gap is something only you can supply: your judgement, your specific knowledge, your voice. Both conclusions are useful, and both mean stopping.
Sometimes the correct outcome is to take the structure and write the rest yourself. That is not a failure of the tool; it is the tool having done the part it was good at.
Two rounds without improvement means the gap is yours to fill, not the model's.