AI image prompts: a structure for controlled, consistent images
AI image prompts are written descriptions that tell an image model what to draw. A controlled prompt names seven things in a fixed order: the subject, the setting, the composition and framing, the light, the style in plain words, the aspect ratio, and what to leave out. The more specific each part is, the less the model has to guess.
Most weak images come from prompts that describe a mood instead of a picture. A fixed structure turns the request into a short brief that can be checked, changed one line at a time and reused for a whole series.
Key ideas
- Write the prompt in a fixed order: subject, setting, framing, light, style, aspect ratio, exclusions.
- Change one variable per attempt, so each new image tells you what that change did.
- Keep a series consistent with a locked base prompt and the same reference images.
- A prompt alone does not make you the author in US copyright terms; label generated images honestly.
The seven parts of a controlled prompt
Google's prompt guide for its Imagen models splits a prompt into subject, context and background, and style. For work that has to match a layout, add four more parts. Composition and framing says how close the camera is and from where: close-up, aerial, from below. Light names the source and quality: natural, warm, cold, dramatic. Aspect ratio fixes the shape before anything else, because a square and a 16:9 frame need different compositions; Imagen offers 1:1, 4:3, 3:4, 16:9 and 9:16.
Exclusions come last. Google advises naming unwanted elements plainly, as in "wall, frame", rather than writing "no" or "don't", which models can misread. If text must appear in the image, Google recommends keeping it to 25 characters or less and to two or three short phrases at most.
Describe style in words, not names
A style reference works best as a description of what you see: the medium (watercolour, matte 3D render, 35mm film photograph), the palette, the lens and the texture. Google's Gemini guide gives the same pattern for photographs, naming the shot type, subject, setting, lighting, camera angle and lens, and states that the more specific you are, the more control you have. A described style can be repeated, explained to a client and changed one word at a time. A borrowed name cannot.
Iterate one variable at a time
Google's guide says iteration is important and that the first result is often not the best. Treat each attempt as a test. Keep the prompt fixed, change one line, generate, and write down what moved. If you change the light and the lens together, you will not know which change fixed the picture.
For edits, describe only the change. Google's Gemini examples ask the model to change only one named element and keep everything else the same. That sentence is often the difference between a small correction and a new image.
Series, rights and labels
To keep a set of images consistent, lock a base prompt that holds the style, light, framing and ratio, and vary only the subject line. Where the model accepts reference images, reuse the same ones for every image: Gemini's image models take several character references for this purpose.
On rights, the US Copyright Office concluded in January 2025 that generated output is protected only where a human author determined sufficient expressive elements, and that the mere provision of prompts is not enough. Human arrangement or real modification of the output can count. On labelling, Google states that all images generated through the Gemini API carry a SynthID watermark; say plainly in captions and client files that an image was generated. The VITON13 SCHOOL course AI for designers covers this whole path: concept, art direction, image generation, brand systems, interfaces and rights. The free prompt brief template helps you write the request down before you type it.
In everyday language
A prompt is a shot list for a photographer who cannot ask questions. Every line you leave out, the model fills in with a guess.
Try it yourself
Open the free prompt brief template, then write one image prompt in the seven-part order for a product photo of a ceramic mug. Generate it, then change only the light line from warm daylight to cold studio light and generate again.
Expected result
Two images with the same subject, setting, framing, style and ratio, where only the light differs. If anything else changed, a line in the base prompt was too vague.
Check your answer: Why is it better to write "text, logo" in the exclusions than "don't add any text or logo"?
Google's guidance is to name unwanted elements plainly. Instructive words like no and don't can be misread, so the model may add the very things the sentence mentions.
Questions
How long should AI image prompts be?
Long enough to cover the seven parts, which is usually two to four sentences. Length is not the goal. A short prompt that names the framing and light controls more than a long one full of adjectives.
Can I sell images made from my prompts?
That depends on the model's terms and on your country's law. In the US, the Copyright Office's view is that prompts alone do not create authorship, so check the terms and add real human work if you need protection.
Write your next prompt in the seven-part order and change one line per attempt. The VITON13 SCHOOL prompt brief template is free; the AI for designers course goes further, into art direction, brand systems and rights.