How to Write a Prompt That Actually Gets the Picture

Most disappointing images come from prompts that were too short, not prompts that were too clumsy. "A cat in a city" leaves every real decision — lens, light, mood, framing — to the model, and the model picks the average of everything it has seen.

The fix isn't longer prose. It's saying the four or five things that actually decide the picture.

The order that works

Subject → medium or camera → light → framing → the things to avoid.

Take the plain version:

Studio portrait of a woman.

Now add each layer:

Studio portrait of a woman in a cream knit sweater by a window, soft daylight, shallow depth of field, natural skin texture.

That's one sentence, and every clause is doing work. "By a window" and "soft daylight" decide the lighting. "Shallow depth of field" is a lens instruction. "Natural skin texture" is the one that stops the plastic-retouch look most models drift into.

Photographic vocabulary works because the models learned from captioned photographs: 85mm, f/1.8, blue hour, backlit, macro, overhead flat lay. So does the vocabulary of any other medium — watercolour, flat vector, isometric 3D render, engraving, film still. Pick a medium explicitly and the whole image follows it.

Four things that usually go wrong

The text comes out as gibberish

Quote the exact words, keep them short, and say where they go: a poster with the headline "Winter Sale" across the top. Then pick the right model — Nano Banana Pro and GPT Image 2 render legible text; FLUX Schnell is a draft model and will approximate it. Long paragraphs are still a gamble on any model.

The composition is wrong

Say the framing rather than hoping for it: centred, rule of thirds, low angle, overhead, full body, tight crop on the hands, room at the top for a headline. Choose the aspect ratio before generating too — 16:9 and 9:16 change what the model composes, not just where it crops.

The style drifts across a set

Write one style clause and reuse it verbatim on every render in the set — same medium, same palette, same lighting words. Changing the subject while keeping the style sentence identical is what makes a page of illustrations look like one commission instead of five.

It ignored half the prompt

That's usually a model choice, not a wording problem. GPT Image 2 is the literal-minded one: counts, placement and labels come back as written. If your prompt is a brief — three objects, this in the corner, that at the bottom — put it there.

Say what you want, not what you don't

"No text, no watermark, not blurry" tends to work less well than describing the positive: a clean, unmarked surface, sharp focus throughout. Negative phrasing puts the concept in the prompt, and some of it leaks into the picture.

Iterate on one variable

When an image is close, change one clause at a time — light, or lens, or medium, never all three. On FLUX Schnell that costs a credit a render and gives you four at once, which makes it the cheapest place to learn what your wording is doing. Once the sentence is right, re-run it on a heavier model.

Steal from working prompts

Every example in the prompt library is a real prompt with the picture it produced. Click one and it loads straight into the box, model preselected — change the subject, keep the structure, and you've started from something that already worked.

Ready to try it? Open the generator.