Making it ask you questions costs less than three failed attempts

The bad draft usually traces back to an unasked question

You ask for a page of copy, receive something reasonable in shape and wrong in almost every particular, and spend the next three attempts correcting assumptions. The audience was wrong. The length was wrong. The thing was pitched as a launch when it was a renewal. None of these were mysteries; they were simply never stated.

The default behaviour is to proceed. Given an incomplete brief, the tool fills the gaps and produces something rather than stopping to ask, which is the right default for most quick requests and the wrong one for any task with an actual specification behind it.

You can change that default with a sentence. Before writing anything, ask what would materially change the output, and wait. The effect is not subtle, because the gaps that get filled silently are exactly the ones that make three attempts necessary.

Ask for the questions before the work

The instruction that works is narrow: list the questions whose answers would change what you produce, ask nothing you can reasonably infer, and do not begin until they are answered. Without the second clause you get a generic questionnaire, including several items that are obvious from what you already wrote.

Limiting the count helps too. Five questions, ranked by how much the answer matters, forces a selection and gives you a list you will actually answer. An unbounded list tends towards thoroughness, which is the same failure as an unbounded draft.

You do not have to answer all of them. Answering three and saying that the rest are your call is a perfectly good reply, and it is more informative than silence because it tells the tool which decisions have been delegated deliberately rather than by omission.

The questions themselves tell you about your brief

Read the list before answering it and treat it as a diagnostic. If the questions are things you thought were obvious from your request, your request was less clear than you believed, and that is worth knowing regardless of what the tool eventually produces.

If the questions are all trivial — preferred heading style, whether to use bullet points — the brief was probably clear enough already and the interview step is not earning its turn. That is a useful signal to skip the step next time on this kind of task.

Occasionally a question exposes something you had not decided at all. Who signs this off, whether the reader has seen the previous version, what happens if the answer is no. Those are questions about the work rather than about the prompt, and they would have surfaced later at greater cost.

Where the interview is a waste of a turn

For small, bounded tasks it is pure overhead. Reformatting a list, converting units, tightening a paragraph you have already written: the specification is contained in the request and asking for clarification produces polite noise. Use the step where the output is long, or where a wrong assumption costs a rewrite.

It is also weak on tasks where the important unknowns are facts rather than preferences. If the real gap is a number nobody has, no question surfaces it, and a list of questions can create a false impression that the brief is now complete when the missing piece was never a matter of preference.

And it can be gamed by the shape of the request itself. A question list generated from a vague prompt tends to be vague, so the technique amplifies whatever clarity you brought rather than substituting for it. It is a multiplier, not a replacement for thinking about what you want.

The answers are the brief, so keep them

What comes out of a short interview is a compact statement of audience, scope, constraints and tone, written in a form that produced good output once. That is precisely the artefact worth saving, and it is far more reusable than the prompt that preceded it.

Fold the answers back into a single briefing block for next time rather than repeating the interview on every similar job. The second run on the same kind of task should start from the accumulated brief and ask only what is genuinely new about this instance.

Over a few months this quietly becomes a house specification: how long things run, who reads them, what is never mentioned, which decisions belong to a person. That document is useful to colleagues who have never typed a prompt in their lives, which is a fair indication that the value was in the specification rather than the tooling.