Ask for options that differ along an axis, or you get one idea five times

Five options and one idea

Ask for five versions of an opening paragraph, a proposal or an explanation and you will usually get five outputs sharing a structure, an emphasis and a length, differing mainly in word choice. They look like a set of options. They are one option rendered five times, and choosing between them is choosing between synonyms.

This isn’t perversity. Nothing in the request said the five should be unlike each other, so each was produced against the same instruction and landed near the same place. Variety wasn’t a requirement, it was an assumption, and assumptions do not survive the trip.

The cost is subtle and it compounds. You believe you have surveyed the space and picked the best of five, when you have actually accepted the first idea and then decorated it. The comparison felt like a decision, which is precisely what makes it hard to notice.

Difference has to be given a direction

The fix is to name the dimension along which the options must differ, and to name it in the request rather than hoping for it. Length. Formality. Which fact leads. Whether the reader is assumed to know the background. Whether the piece is a warning, an offer or a description.

Once the dimension is stated the outputs separate immediately, because now each version has a distinct job. One leads on the price, one on the timing, one on who it is for. Those are genuinely different pitches, and it is possible to disagree about which is right.

A more demanding version of this works well: ask for options that a reasonable colleague would argue about. That framing rules out the safe middle by construction, since two versions nobody could argue between aren’t really two versions.

Generate them apart, then compare them together

Options produced in one pass tend to converge, because everything already on the page influences what comes next. The second suggestion is written in the shadow of the first, and by the fourth the pattern has hardened into a template that the fifth simply completes.

Producing them in separate runs, with the same brief and no sight of the earlier attempts, gives real spread. It costs more and is worth it whenever the decision matters, which is a narrower set of cases than people assume — most of the time a near-identical set is perfectly adequate.

The comparison, though, should happen in one place, side by side, with the axis you asked for made visible. Reading options sequentially favours the last one you saw. Reading them in a column, with a label on what each one is trying to do, favours the one that fits.

One caution about separate runs: they only spread if the brief itself leaves room. A brief that already names the audience, the length, the leading fact and the register has decided nearly everything, and running it three times in isolation will still return the same piece three times over.

Rejects are information if you write down why

The value of a spread of options is frequently not the winner. It is the moment you look at the third one and realise the whole framing is wrong, or that a constraint you had not articulated has just been violated in an instructive way.

So record the reason for rejecting each, in a few words. Too corporate. Assumes they already know us. Overpromises. That list is a specification, arrived at backwards, and it is usually better than the specification you would have written from a blank page.

Feeding those reasons back into the next round is more effective than asking for another five. You have converted a vague preference into stated criteria, and stated criteria are the thing that makes a second round better rather than merely different.

When a single considered draft is the better ask

What is described here is a small deliberate spread for a decision about approach: how to frame a document, which argument leads, what structure the thing takes. That is a different technique from mass-generating candidates for a headline or a name, where the volume itself does the work and selection is nearly free.

They suit long or technical writing much less. Five versions of a two-thousand-word explanation cannot be compared properly, so you will skim them, pick the one with the best opening, and discard four pieces of work you never read. That is not selection, it is sampling with extra steps.

And where there is a correct answer rather than a preferable one, asking for alternatives is actively misleading. A set of five plausible interpretations of an ambiguous clause invites you to choose the one you like, when the job is to find out which is right — a job that happens outside the tool entirely.