A prompt rewriter that proposes settings too

Almost every image tool now has an "improve this prompt" button. Most of them do one thing: expand the sentence. Add adjectives, add a lighting clause, add 8k, highly detailed, trending on artstation if they were built in 2023.

That is the least useful half of the job, because the settings are where new users actually lose.

The settings are the part nobody picks

Watch someone use a generator for the first time. They type a sentence, they press the button, and they touch none of the controls — because the controls are named for the API rather than for what they are about to make. Nobody knows whether a pitch deck slide wants 16:9 or 3:2, or whether text-heavy work needs a quality tier above the default.

So the prompt itself is the best available signal of what the settings should be. A pitch deck slide titled… is a 16:9 request whether or not the user knows it. A product shot for a listing is square. A poster is portrait and probably wants a higher tier, because posters are mostly typography and typography is where the cheap tiers fail first.

Propose, do not overwrite

The rule that keeps this from being annoying: apply the suggested settings only if the user has not touched the controls.

const untouched =
  JSON.stringify(current) === JSON.stringify(defaults);
if (!untouched) return current;   // they chose; leave it alone

Overwriting a tier somebody deliberately set is the kind of help nobody asked for. Untouched defaults are a different thing — those are ours, and the entire point of computing settings is that a computed one beats a default.

Two failure modes to design against

The rewriter makes the prompt worse. Long expansions drift. If the user typed something specific and got back something generic-but-longer, that is a regression even if it scores better on a rubric. Show both and let them keep theirs.

The rewrite blows the length ceiling. Mode instructions and constraint clauses get added server-side, after the user stopped typing. A rewrite that fits on its own can still produce a composed prompt that is over the limit, and the error arrives talking about a length the user never saw. Count the composed length, not the typed one.

The honest framing

A rewriter is not a quality multiplier. It is a way of transferring the conventions a regular user has not learned yet — which frame, which tier, which words the model responds to — and its value drops sharply once somebody has made fifty images. Build it for the first session, and do not be offended when power users turn it off.

Ours is free and does the settings half: gptimage25.top/prompt-generator.