How to write prompts for AI image generators
An AI image generator turns a text description into a picture. It cannot read your mind, so the words you choose decide most of the result. This guide shows a prompt structure, real example prompts and a way to revise them.
A six-part prompt structure
You do not need every part every time, but running through the list stops you leaving out something that matters.
- Subject: what is in the picture. “A ceramic teapot”.
- Setting: where it is. “On a wooden kitchen table by a window”.
- Style or medium: “watercolour illustration”, “flat vector graphic”, “35mm photograph”.
- Composition: framing and angle. “Close-up, eye level, subject on the left third”.
- Light and colour: “soft morning light, muted greens and cream”.
- Format: the shape you need, such as square or wide, if the tool lets you set it.
Example prompts, from vague to specific
Vague: a cat
Better: a ginger cat asleep on a stack of books, cosy reading nook
Specific: a ginger cat asleep on a stack of old hardback books in a cosy reading nook, warm lamp light from the right, shallow depth of field, photograph, muted browns and oranges, wide format
The specific version names the light direction, the depth of field and a colour palette. Each of those removes a guess the model would otherwise make for you.
More examples for different jobs:
- Blog header:
flat vector illustration of a laptop, a coffee cup and a plant on a desk, top-down view, pastel blue and peach palette, plenty of empty space on the right for text, wide format - Abstract background:
abstract fluid acrylic texture, cells of turquoise and gold on deep navy, high detail, no text - Concept sketch:
pencil sketch of a small timber cabin on stilts above a lake, side elevation, loose lines, white paper background
Revising a prompt step by step
Change one thing at a time so you learn what each word does.
- Generate a few images from your first prompt.
- Write down what is wrong in plain words: “too dark”, “cat looks cartoonish”.
- Add or swap one phrase that addresses the biggest problem.
- Generate again and compare with the previous batch.
- Keep a note of the final prompt next to the image you used.
What these tools often get wrong
- Text inside images may come out misspelt. Add text afterwards in a design tool.
- Hands, counts and small details can be off. Check closely before you publish.
- Negations such as “no people” are not always respected; some tools have a separate field for things to exclude.
- Precise layouts are hard. For a strict layout, generate parts and arrange them yourself.
Rights and honesty
Read the terms of the tool you use for what you may do with outputs, including commercial use. Avoid prompts that copy a living artist’s name or a trademarked character. Where your audience would care, say that an image was made with AI.
Questions
Do longer prompts always work better?
No. Extra words can pull the image in directions you did not intend. Add detail where the results go wrong, and stop when they look right.
Will the same prompt give the same image?
Usually not. Most tools add randomness, so each run differs. Some let you fix a seed value to repeat a result.