PTN—20

Variations

For subjective output, generate labeled alternatives instead of forcing serial regenerate-and-pray.

When a task has one right answer, generate it. When it has taste — a headline, a layout, an opening paragraph — a single output is a coin flip presented as a conclusion. The regenerate button is the wrong tool for that problem: it discards a candidate to produce another blind draw, one at a time, with no memory and no map. Variations replace the slot machine with a spread: several alternatives at once, side by side, so the person judges by comparison instead of by sequential gambling.

The labels are what make a spread navigable. “Option A / B / C” tells the person nothing except that there are three; labels that name what differs — “Shorter”, “Leads with the data”, “Warmer open” — turn the set into a legible tour of the output space. That legibility feeds back into expression: someone who picks “Leads with the data” has just told the system something no rating widget would have captured, and the next spread can vary along the axis they cared about.

Regeneration inside this pattern must respect the stated constraints. If the person fixed the length and asked for tone alternatives, a regenerate that also rewrites the length has broken the contract — vary what was left open, hold what was pinned. And choosing a variant should be a promotion, not an execution: the chosen one takes the stage, the others recede but stay recoverable for the session, because taste is fickle and the discarded option is often the one someone wants back ten minutes later.

The economics are honest to state: a spread of four costs four generations, and not every request earns it. Variations belong where output is subjective and judgment is cheap — glanceable text, images, layouts — not where evaluating each candidate takes real effort. A spread of four twenty-page reports is not choice; it is homework.

Anatomy

Fig. 01 — VariationsLive specimen

Reply to a 2-star review · three takes

Direct · two sentences

You’re right — Marcus arrived 90 minutes late and that’s on us. We’d like to send him back this week, no charge, to finish the tune-up properly.

Warm · leads with the apology

We’re sorry, Mrs. Okafor — waiting 90 minutes past the window is exactly what we promise not to do. Marcus feels the same, and we’d like to make it right with a free return visit this week.

Formal · cites the visit record

Our dispatch log confirms your 8–10 am window was missed; Marcus arrived at 11:28. We apologize for the disruption and would like to schedule a complimentary return visit at your convenience.

All three name the tech and offer the make-good visit.

A spread of alternatives labeled by what differs, with one promoted and the rest recoverable.

1Difference labels. Variants are labeled by what actually differs — tone, length, structure — never “Option A/B/C”.
2Constraint-preserving regenerate. Regenerating varies what was left open and holds what the person pinned.
3Promotion, not deletion. Choosing a variant promotes it to the working copy without destroying the others.
4Recoverable discards. Passed-over variants stay reachable for the session — taste changes its mind.

When to use

The output is subjective — tone, style, composition — and “best” is a judgment only the person can make.
People are hammering regenerate serially, hoping the next blind draw lands.
Alternatives can be compared at a glance, so a spread costs seconds to evaluate, not minutes.
Early exploration, where seeing the space matters more than perfecting one point in it.

Design considerations

01Generate variants that genuinely differ along nameable axes; three near-duplicates is cost without choice.
02Label each variant by its distinguishing quality, in the user’s vocabulary.
03Hold stated constraints constant across the spread and across regenerations.
04Keep spreads small — three or four — and comparable in a single view without scrolling.
05Promote the chosen variant and keep the rest recoverable until the session ends.
06Use the choice as signal: the axis someone picked on is the axis to vary next.

Pitfalls

✕Fake variety — four variants that are the same answer in slightly different clothes, spending 4x compute on 1x choice.
✕Anonymous options: “A/B/C” labels that force the person to diff the variants by hand.
✕Destructive selection, where picking one deletes the others and second thoughts have nowhere to go.
✕Spreading everything — variants on factual answers, long documents, or anything expensive to evaluate.

In the wild

Midjourney: Every prompt returns a four-image grid, with per-image variation controls that explore around the one you liked — comparison is the default, not an upgrade.
GitHub Copilot: Beyond the inline ghost text, a completions panel lists alternative suggestions so developers pick between candidates instead of rerolling one.
Adobe Firefly: Generations arrive as a set of variations per prompt, and Generative Fill offers multiple candidate fills to flip between in place.