How to write image prompts: one brief, three ways, four renders

Write an image prompt the way you would brief a photographer. Name the subject, the setting, the light, the lens and framing, the style, and what must stay out of the shot. We wrote one lighthouse brief at three lengths and rendered each on GPT-Image-2.5 Flare. Only the full brief came back as the picture we had in mind, and the four-word version came back as a postcard the model composed for us. A prompt is a brief, not a spell, and every part you leave out is a decision you hand to the model.

We ran all four prompts through OpenAI's Image API at the same settings: a 1024×1024 square, medium quality, PNG, one image each. Every image below was generated by GPT-Image-2.5 Flare, and each is a single draw for its prompt.

What goes into an image prompt?

Six parts carry the result: subject, setting, light, lens and framing, style, and exclusions. The makers' own guides land on much the same list. OpenAI's prompting guide says to start with the image you need, then describe subject, composition, style and constraints. Google's Nano Banana guide asks about style, subject, setting, action and composition. Amazon's Nova Canvas guide lists subject, environment, pose, lighting, camera position and style.

The guides disagree on how much the extra detail is worth. Google says even short prompts make high-quality images. Runway says simple prompts do well and detailed ones give finer control over each element. So detail buys control, not quality. A short prompt makes a good image. It just may not make yours.

Here is the brief at three lengths. The terse one names only the subject and the time of day:

A lighthouse at dusk.

The medium one adds the subject's look, the setting and the light:

A red-and-white lighthouse on a rocky headland at dusk, waves breaking below, warm light in the lamp.

The full one adds framing, lens, colour treatment and three exclusions, with one sentence for each part:

A red-and-white striped lighthouse on a rocky headland, centered in the frame. Dusk over a rough sea, waves breaking white on the rocks below. The last orange glow on the horizon and the lamp just lit. Shot on a 35mm lens from a low angle, the foreground rocks sharp. Muted, film-like color. No people, no boats, no text.

What did each version change?

The terse prompt got a competent image, and the model made almost every choice in it. It painted the tower plain white, added a red-roofed keeper's cottage and a fence, lit the lamp, set the sun on the horizon and lit the sky in saturated orange and violet. That is the model's default lighthouse, and it may do as stock filler. A brief with a palette or a layout leaves no room for a coin toss you did not know you were making.

Naming three more things got three more things. The tower turned red and white and stood on a rocky headland, waves broke into spray below it, and the lamp glowed warm. Everything we did not name stayed the model's call. The cottage and fence remain, the sunset is still saturated, and the tower still sits to the left with the sea running off to the right.

The full brief is the first render that looks directed. The striped tower stands centred on a dark outcrop, the camera sits low behind sharp foreground rocks, the surf breaks white, and the sky has gone overcast with only a band of orange at the horizon. The colour is muted as asked, and no people, boats or lettering appear. The cottage, which none of the prompts mentioned, is gone too. Read "35mm" as a request for a look rather than a focal length you can measure. OpenAI's guide calls camera specifications cues for appearance, not a promise of exact physical simulation.

Does the order of the parts matter?

In this pair, the reversed brief kept the content and came back with different framing. The reordered prompt uses the full brief's six sentences in reverse, so the lighthouse arrives last:

Muted, film-like color. No people, no boats, no text. Shot on a 35mm lens from a low angle, the foreground rocks sharp. The last orange glow on the horizon and the lamp just lit. Dusk over a rough sea, waves breaking white on the rocks below. A red-and-white striped lighthouse on a rocky headland, centered in the frame.

Every instruction still landed. The render has the striped tower, the lit lamp, white surf, an orange horizon and muted colour, and none of the excluded things appear. What differs is the framing. The tower is a small figure on a distant island, a low building sits at its base, and the foreground rocks take the bottom half of the frame. These two renders show that difference, but with one draw per prompt they cannot show whether the order or chance caused it.

The makers' guides do weigh in on order. Amazon says to put the least important details near the end of a long prompt. Black Forest Labs' FLUX guide says to front-load what you want when the model keeps missing it. OpenAI suggests organising complex requests as scene, subject, details and constraints.

How do you say what must not appear?

On Flare, in these two renders, saying it plainly worked. The full brief ended by ruling out people, boats and text, and neither render that carried that line showed any of them. OpenAI's guide tells you to state exclusions such as unwanted text, logos and watermarks outright.

That advice does not carry over to every model. Runway says negative phrasing is not supported in Gen-4 Image and may produce the opposite. Black Forest Labs says FLUX models do not support negative prompts and tend to fix on the very word you meant to rule out. Amazon says Nova Canvas does not understand negation, so a fruit basket with no bananas gets bananas, and it gives exclusions their own negative-prompt parameter instead. On those models, describe what fills the space: BFL suggests asking for clean, unmarked surfaces rather than "no text". Write the exclusions for the model you are on, not the one you used last.

Does a longer brief cost more?

In these four renders, not by any amount you would notice. All four returned the same number of image output tokens, and our estimate of what each cost at list price is about 1.3 cents. The Flare model page lists text input, which is how prompt words are billed, at $5 per million tokens. OpenAI's image generation guide notes that token use can differ by model and quality setting, so measure your own prompts. Check the model page for current rates before you budget.

What should you try next?

Write your next prompt as six short sentences, one per part, with the subject first and the exclusions phrased for your model. Render it, change one sentence, and render again. OpenAI's guide advises refining one thing at a time and repeating requests to see how consistent a result is. The model will always fill the gaps you leave; the craft is choosing which ones.

Frequently asked questions

What should an image prompt include?

Name the subject, setting, light, framing and lens, style, and what must not appear. The prompting guides from OpenAI, Google, Amazon and Runway share subject, setting, composition and style; OpenAI also tells you to state exclusions, and Amazon gives them a separate negative prompt.

Does word order matter in an image prompt?

In our pair, the reversed full brief kept every element but came back with a smaller lighthouse and the foreground rocks filling the frame, though one draw each cannot rule out chance. Amazon advises putting the least important details last, and Black Forest Labs advises front-loading what you want.

Should I write "no text" in an image prompt?

On GPT-Image-2.5 Flare, our two renders whose prompts ruled out people, boats and text showed none of them. Runway, FLUX and Nova Canvas say negation can backfire, so on those models describe what fills the space, or use Nova's negative-prompt parameter.

Do longer image prompts cost more?

Barely, in our four renders: each came to an estimated 1.3 cents at list price, from an 11-token prompt up to an 84-token one. The Flare model page lists text input at $5 per million tokens.

Is a short prompt ever enough?

Yes, when any good image will do: our four-word prompt returned a polished lighthouse scene, just not one we had chosen. Once the brief has a palette, a layout or things to keep out, write the full version.