The anatomy of a prompt that works
Four parts, in a deliberate order. Most weak results are a missing part rather than a missing keyword.
Four parts, in order
A prompt is not a search query. You are describing a photograph or painting that does not exist yet, and the model fills every gap you leave. The reliable structure is subject, then action, then setting, then treatment.
- Subject — who or what the frame is about, with the two or three details that make it specific.
- Action or pose — what they are doing, and where they are looking.
- Setting and light — where this happens and what is lighting it.
- Treatment — medium, lens, palette, and the style you want it rendered in.
A silver-haired sorceress in a frost-rimed cloak, turning to look back over her shoulder, standing on a frozen lake at dusk, low sun behind her throwing long shadows across the ice, painterly digital illustration, cool blue palette with a single warm rim light
Each clause answers one of the four parts. Nothing is decorative.
Front-load what matters
Words near the start of a prompt carry more weight than words near the end. If the subject appears in the twelfth clause, after a long preamble about lighting and lens, expect the subject to drift. Put the thing you would be most annoyed to lose first.
This also means long prompts are not automatically better. Past roughly sixty words most models start averaging your requests together instead of honouring each one. If a detail is not load-bearing, cut it and spend the attention elsewhere.
Describe light, not quality
Words like masterpiece, 8k and highly detailed were useful on older models trained on tagged image boards. Current models are trained on captions written in ordinary language, so those tokens mostly add noise. Lighting is the substitute that actually changes the image.
- Instead of cinematic, say low key with a single hard light from camera left.
- Instead of highly detailed, name the texture you want: wet cobblestone, brushed steel, loose watercolour bleed.
- Instead of professional photo, name the lens: shot on 85mm at f/1.8, shallow depth of field.
Worth remembering
- Cover subject, action, setting and treatment — a missing part is a gap the model fills for you.
- Put the most important detail first; attention thins out toward the end.
- Replace quality adjectives with concrete descriptions of light, texture and lens.