Prompt anatomy: four layers
AI image generation · Lesson 2 / 20
A formula that works
A reliable prompt describes four things: subject, style, composition, and light and mood. Omitting any layer means the model chooses it — and usually not the way you need.
[SUBJECT] a middle-aged woman astronaut, a tired smile
[STYLE] 1960s retro poster, grainy print, limited palette
[COMPOSITION] waist-up portrait, looking at camera, subject on the
right, empty space on the left for text
[LIGHT AND MOOD] warm side light, an air of restrained optimism
Order matters
The closer a word is to the beginning, the more it influences the result. Start with the main subject rather than the style: a prompt opening with "photorealistic, 8k, cinematic" often produces a beautiful picture of the wrong thing.
Specifics instead of adjectives
"A beautiful sunset" is an empty description — the model has its own idea of beautiful. "Low sun, long shadows, orange-to-violet gradient, silhouettes in the foreground" is a description you can draw.
The rule is simple: if an adjective can be replaced by something visible in the frame, replace it.
Length
A long prompt isn't a good prompt. Past a certain volume the model starts losing some of your instructions, and there's no predicting which. Twenty precise words beat eighty where half are synonyms for quality: "masterpiece, high resolution, stunning, professional".
When there are many elements
Models assemble complex multi-object scenes badly: they muddle what's where and lose items. It's easier to generate a simple scene and finish it in an editor than to force a complex composition through the prompt.
Cheat sheet
- Four layers: subject, style, composition, light and mood.
- The main thing goes at the start.
- Replace adjectives with the visible.
- Complex scenes are easier assembled in an editor.
Build a four-layer prompt for a real task of yours. Then make three versions: (1) complete; (2) without the composition layer; (3) without the light and mood layer. Generate all three. Separately, in the complete version replace every evaluative adjective with a description of the visible and generate again. Paste the prompts, describe what changed when each layer was missing, and what the adjective replacement produced.