How to Write AI Image Prompts That Work: A Practical Guide

“A cool picture of a lighthouse” gives a model almost nothing to work with, so it fills the gaps with the most average version of a lighthouse it knows. Professional photographers and illustrators never get briefs like that. They hear who or what, where, when, in what light, shot how, and for what purpose.

That's the whole trick of prompt writing: describe the picture the way you'd brief a person who can't ask you questions. Concrete nouns beat adjectives. “Stunning” and “beautiful” carry no information; “low winter sun behind the lighthouse, long shadows across wet rocks” does.

1. Subject and action. Who or what, doing what. “A 60-year-old fisherman mending a net” beats “an old man”.

2. Setting. Place, time of day, season, weather. “On a harbor wall at dawn, light drizzle.”

3. Light. Direction, quality and color. “Soft overcast light from the sea side, cool blue tones.”

4. Camera or medium. “Shot on a 35mm lens at eye level, shallow depth of field” for photos; “gouache illustration with visible brushstrokes” for art.

5. Style and mood. A genre, era or feeling: “quiet documentary style, muted palette.”

6. Constraints. What must or mustn't appear: “no text, no other people, vertical 4:5 composition.”

Put together: “A 60-year-old fisherman mending a net on a harbor wall at dawn, light drizzle, soft overcast light from the sea, cool blue tones, 35mm lens at eye level, shallow depth of field, quiet documentary style, no text.” You don't need every part every time, but each one you skip is a decision the model makes for you.

Most modern image models in the Image Lab can edit from reference images, and editing has its own grammar: say what to keep before what to change.

“Keep the woman's face, pose and the red coat exactly as in the reference. Change the background to a snowy Montreal street at night with warm shop windows.”

One change per pass. Ask for the background in one generation and the lighting in the next. Several changes at once make it harder to see which instruction the model ignored.

Name the references. When you upload more than one image, say what each one is for: the face from the first, the jacket from the second, the color palette from the third.

Nano Banana 2.1 (default, 12 credits at 2K). Responds well to plain, conversational sentences and to multi-image edits with up to 14 references. Describe relationships between references in words.

GPT Image 2.5 (from 4 credits at 1K low quality). The pick for images with words in them. Put exact text in quotation marks, then describe the typography: “The poster headline reads ‘Night Market’ in bold condensed sans-serif, centered, white on deep red.” It also follows long, layout-heavy instructions closely.

Spicy Imagine Pro (8 credits at 2K). Built on Seedream 5.0 for edits that preserve faces, lighting and color tones from up to 10 references; refer to them as Figure 1, Figure 2 and so on.

The same prompt can look very different across models, which is useful: run your best prompt on two of them and keep the one that reads it best. Our AI image generators comparison covers how they differ.

It looks generic. Add specifics in the setting and light, or a camera or medium. Generic results usually come from generic prompts.

It ignored part of the prompt. Move the important element to the start, cut competing details, or split the job into a generation and an edit.

Contradictions. “Bright and moody, minimal and busy” makes the model pick one at random. Decide.

Unwanted extras. These models don't have a separate negative-prompt box, so state exclusions in the prompt itself: “no text, no watermark, no other people.”

Faces drift between images. Use the same reference image every time instead of re-describing the face.

You don't want to write prompts at all. Templates start from finished looks. After you generate with one, its prompt unlocks, so you can read how a working prompt is built and borrow its style words.