Ask for three, then pick: you can't describe what you haven't seen

Overview

Why now. On 9 September a magazine piece asked working writers how they use AI, and the answers split into two loops. One writer hands over a paragraph he is unhappy with, asks for a fix, then explains, exchange after exchange, what is wrong with each attempt. Another hands over a sentence she cannot get right, asks for several clearer versions to get nearer what she means. Both are working out what they meant; only one has to describe it first.

That is the difference. The standard loop — one answer, then a paragraph on the fix — works when you know what you want. When you don’t, the paragraph is where the time goes: you are specifying something you have not seen. People are far better at recognising than at specifying.

So the habit: when you can’t yet say what you want, ask for three clearly different versions in one go and point at the one that’s closest, before you ask for any rewrite.

The content

The obvious advice is to be more specific. On an open task — a message with an awkward tone, a section that isn’t landing — specificity is exactly what you lack. One draft hands you a specification problem: what, precisely, is wrong with this? Three drafts hand you a recognition problem: which is nearest? The second is the easier question, and the studies point the same way, before language models and since.

In a 2010 Stanford experiment, novice designers made web banner ads either one at a time with feedback after each, or several before any feedback. The parallel group’s ads did better on click-through data and expert ratings, and were more diverse. Iteration, they noted, can also produce fixation.

Now the models. A CHI 2023 study gave 129 people a writing tool offering one suggestion or three, with or without a box for instructions to the AI. People preferred choosing from the three over steering with instructions; single suggestions were rated more distracting, less helpful and more in need of manual editing. An earlier study by the same group had 156 people write emails with zero, one, three or six suggestions: three or six raised the chance of accepting one; the median verdict on three was “just right” — 62% said so — while six started costing time and was rated “too high”.

The one language-model study with a rated quality outcome is the most careful. In Science Advances in 2024, 293 writers were randomised to no AI ideas, one, or up to five; separate readers rated their stories. Stories written with access to up to five ideas were rated 5.1% more useful than stories written with one. Two caveats travel with it: the five arrived one request at a time, not in a single pass, and both AI conditions made the stories more like each other. The gain is real. So is the anchoring.

The other half of the case is the loop you are replacing. A 2025 study ran twelve-turn loops — vague follow-ups like “this is good, make it better”, and targeted ones — against four models on ideation, code and maths. On code, the strongest model was right at the first turn, then decayed to near nothing within a few turns while the code ballooned. The authors’ advice: “if a correct path does not appear quickly, stop or restart, do not push vague refinement.” Their maths result went the other way — iteration helped, most reliably under a targeted prompt to elaborate — which is this module’s scope in one line: vague feedback on open tasks goes nowhere; targeted feedback works; the pick is what makes yours targeted.

Two conditions carry the habit; the folk version drops both.

Three, not many. A large menu makes choosing its own job, and you are using this precisely because you can’t yet say what you want. Three is enough to see the axis you are choosing along; it is the count the email writers rated “just right”.

Clearly different, in the ask. Hitting regenerate does not give you three. A NeurIPS 2025 paper sampled fifty answers to each of a hundred open questions across dozens of models: in 79% of cases the average similarity between answers exceeded 0.8. Regenerating gives you the same answer in three coats. Name the axis — tone, structure, opening, length — and you get a menu.

One honest note: no study runs the exact comparison — three in one pass against one then iterated, on a rated outcome. Each leg rests on an adjacent design; hence “the studies point the same way”, not “a study showed”.

Try it

Take something you asked for this week and then corrected three times.

1. Ask again, differently. “Give me three clearly different versions of this. Vary the [opening / structure / tone] — I’ll tell you which is closest.” One axis is enough.
2. Point, don’t describe. “Two is closest. Keep its opening; take the shorter second paragraph from three.” That is a rewrite instruction you could not have written cold.
3. Then correct once. Compare the message count.

It runs in any chat box your workspace has enabled.

Where it breaks. If you can already say what is wrong, say it; three is then a detour. It is not for tasks with a right answer; the maths result above improved most reliably with targeted iteration. And all three come from one model’s sense of a good answer: the 2024 writers who used AI ideas wrote more useful stories that were also more alike. Three is a wider view of the model’s defaults, not an escape from them; when the axis that matters is yours — your argument, your voice — write the first line yourself and ask for three continuations.

Additional reading

• What Writers Who Use AI Want You to Know — Every — the 9 September piece; both loops.
• Choice Over Control — CHI 2023 — 129 people, one vs three suggestions.
• Multiple Parallel Phrase Suggestions — CHI 2021 — 156 email writers; 62% rated three “just right”; no quality measure.
• Generative AI enhances individual creativity but reduces the collective diversity — Science Advances 2024 — 293 writers, pre-registered; the 5.1% gap and the homogenisation caveat.
• Parallel Prototyping — ACM ToCHI 2010 — the pre-AI origin.
• Another Turn, Better Output? — arXiv:2509.06770 — twelve-turn loops; code collapsed under vague prompts, maths improved under targeted ones.
• Artificial Hivemind — NeurIPS 2025 — why “regenerate” is not “three different”.

Editor’s note

I do this almost every time I encounter a problem that fits this pattern. I don’t always ask for three options, sometimes I feel like I only need two, and other times I want many more when each option is only a line long. But the logic is the same. What I often find is that one option is clearly better than the others once I can see them side by side (often the first option I’m given), which still helps me to identify the direction I want to go.