How to write AI image prompts that work

Published: June 16, 2026

A good AI image prompt is a short, concrete description that names five things: the subject, the setting, the lighting, the framing, and the style. Most disappointing results come from leaving three of those five to the model to guess. Once you describe all five in plain language, the output becomes far more predictable — and far closer to what you pictured.

The five parts of a usable prompt

Think of a prompt as a sentence you would say to a photographer or illustrator who cannot see what is in your head. Cover each part briefly and in order:

• Subject — what the image is of: "a black ceramic coffee mug", "a mid-century living room", "a two-storey house".
• Setting — where it sits: "on a pale oak table", "with a large arched window", "surrounded by tall pines".
• Lighting — how it is lit: "soft side light", "warm afternoon sun", "even studio light".
• Framing — the shot: "hero product shot, straight on", "wide angle", "top-down flat-lay".
• Style — the look: "clean catalogue photography", "watercolour", "cinematic, muted palette".

You do not need every part for every image, but naming each one removes a decision the model would otherwise make at random.

From vague to specific: a worked example

Start with a weak prompt: "a nice photo of a chair". The model has to invent the material, the room, the light, and the angle, so every generation looks different. Now add the five parts: "a walnut mid-century armchair with mustard upholstery, in a sunlit living room with a large window, warm afternoon light, photographed straight on at eye level, clean interior-photography style." The second prompt returns a consistent, on-brief image because nothing important was left to chance.

Describe what you want, not what you do not

Models respond best to positive, present descriptions. "An empty, uncluttered desk" works better than "a desk with no mess". Spend your words on the thing you want to see rather than listing everything to avoid. If something keeps appearing that you do not want, change the surrounding description so there is no room for it, rather than adding a long list of exclusions.

Use a reference image to lock the subject

When the subject must stay accurate — your actual product, a specific room, a real façade — upload a reference image instead of trying to describe it perfectly in words. With a reference, the prompt only has to handle the parts you are changing: the background, the lighting, or the style. This is the single biggest accuracy improvement available, and it turns a long, fragile prompt into a short, reliable one.

Iterate, because it is cheap

Prompt writing is not a one-shot skill; it is a fast feedback loop. Generate, look at what the model got wrong, change the one part of the prompt responsible, and generate again. Because each render takes seconds and costs only a few credits — and failed jobs are refunded — the right habit is to make several small adjustments rather than agonising over a perfect first prompt. After a handful of iterations you will have both the image and a reusable prompt you can apply to the next one.

A reusable checklist

• Name the subject, setting, lighting, framing, and style.
• Write in positive terms — what you want to see.
• Upload a reference when the subject must stay accurate.
• Keep it concrete; drop adjectives that do not change the picture.
• Generate a few variants and keep the best.