Write a useful image prompt
Image prompts fail differently from text prompts. A vague text prompt produces generic writing you can edit; a vague image prompt produces an image you have to pay to replace.
Start with the style preset
Before writing anything, pick the right preset. It carries more of the result than several sentences of prompt would:
| Preset | Produces |
|---|---|
| Realistic portrait | A photographic image of a person |
| Product photography | A studio-style product shot |
| Realistic landscape | A photographic scene or environment |
| Web illustration | A flat, graphic illustration |
| 3D illustration | A rendered, dimensional illustration |
| Marvel Comic style | A comic-book treatment |
| None | Only what your prompt says |
Choose None when you want full control and are prepared to describe the medium yourself.
What a prompt should contain
Work through four things, in this order:
1. The subject. What is in the picture, specifically. "A woman working" is weak; "a technician inspecting a circuit board with a magnifier" is a picture.
2. The setting. Where it is. A background left unspecified will be invented, often as a generic office.
3. The light and mood. "Soft morning light through a window", "hard studio light on a dark background", "overcast and flat". This does more for realism than any other single element.
4. The composition. Close-up or wide, where the subject sits, whether there is empty space. If the image needs room for a headline, say so: "wide shot with empty space on the left".
The prefilled example on the screen shows the expected level of detail:
An image of a gorgeous landscape with mountains behind, using warm and light blue tones.
That is about the right length. Two or three sentences.
What to leave out
Text. Generated lettering is unreliable — misspelt words, invented characters, broken kerning. If the image needs words, generate it without them and add the text in your page builder or an image editor.
Real people. Do not ask for a named person or a recognisable likeness, and do not use a generated face where a real photograph is implied.
Brands and logos. Trademarks that drift into an image are a problem you will have to solve later.
Long lists of adjectives. "Beautiful, stunning, amazing, professional, high-quality, 8K, award-winning" adds nothing beyond the first term. Describe the picture, not your hopes for it.
Contradictions. "Minimalist but richly detailed" resolves arbitrarily.
Colour and brand
POM AI automatically supplies the relevant visual-profile sections, including palette, materials, composition and exclusions. The individual prompt should describe this image's subject, setting and job rather than repeating the permanent profile.
If one image deliberately departs from the normal art direction, say so in the prompt. The task-specific instruction should be precise enough to explain the exception without rewriting the entire visual identity. See Provide useful business context.
Iterating without wasting credits
- Start with Nano Banana 2 (Google Gemini) at 1K, or estimate one image with your preferred model.
- Change one element per attempt — subject, then light, then composition.
- When the composition is right, compare a different model, resolution or OpenAI quality if the result needs improvement.
- Save the winning prompt as a template.
Changing three things at once tells you nothing about which one worked.
Prompts for common needs
Hero band
Wide interior of a modern workshop, soft daylight from the left, machinery slightly out of focus in the background, generous empty space on the right for text.
Product shot
Single ceramic mug on a pale grey surface, clean studio lighting, subtle shadow, no props, shot slightly above eye level.
Blog illustration
Flat illustration of a simple network of connected nodes, muted blue and sand palette, plenty of white space, no text.
Abstract background
Soft gradient texture in warm neutral tones, very low contrast, no focal point, suitable as a background behind white text.
That last one shows a useful habit: describing the job the image has to do, not only what it looks like.
Editing prompts are different
The improve and background-removal tools also take prompts, but they modify an image you supply rather than inventing one. Their prompts describe a change — "make the background a rich wooden surface" — not a whole scene. See Improve an existing image.
When the result is close but wrong
If two or three attempts are close but not right, consider whether the picture you want exists as a photograph you could take or license. Generation is good at plausible generic imagery and poor at a specific thing that must be exactly right.