How to Write AI Image Prompts That Actually Work
Most prompts fail for the same three reasons. Here is the structure professional prompt writers use — subject, setting, light, lens, mood — and why order matters more than length.
A prompt is a brief. If you handed the same sentence to a photographer, would they know what to shoot? That single question fixes most bad prompts before you ever hit generate.
The five-part structure
Almost every prompt that works follows the same order. Models read the beginning of your prompt with more weight than the end, so the sequence is not cosmetic.
- Subject — who or what, with one distinguishing detail.
- Setting — where they are, and what is behind them.
- Light — the source, its direction, and its quality.
- Camera — focal length and aperture, which control compression and depth.
- Mood and finish — palette, grade, level of detail.
Write those five in order and you have covered ninety percent of what a model needs. Everything after that is refinement.
Describe instruments, not adjectives
"Beautiful lighting" means nothing. "A large softbox at forty-five degrees camera left with a black flag on the shadow side" means one specific thing, and the model can render it. The most reliable upgrade you can make to any prompt is replacing an adjective with a piece of equipment.
The same applies to camera language. "Close-up" is vague; "100mm macro at f/8, focus stacked" produces a predictable result every time.
Realism lives in imperfection
AI images look synthetic when they are too clean. Ask for the things a real camera cannot avoid:
- Visible skin pores and fine flyaway hairs on portraits.
- Dust in the light beam, micro-scratches on metal, condensation on glass.
- Slight asymmetry — a crooked collar, an uneven stack, a fingerprint.
The negative prompt is half the work
A negative prompt is not a dumping ground for every word you dislike. Keep it to the specific failure modes of the shot you are making. A portrait needs plastic skin, extra fingers, deformed eyes. A product shot needs visible support stand, cluttered background, colour fringing. Copying a generic 200-word negative list into every prompt actively degrades your results.
Iterate on one variable at a time
When a generation is close but wrong, change one thing. Move the light. Change the lens. Adjust the palette. If you rewrite the whole prompt you learn nothing, and you cannot repeat the result on the next project.
Save what works
The people who get consistently good output are not writing better sentences on the fly — they are reusing a library. Keep the prompts that work, note the model and settings you used, and treat them as presets. That is exactly what this library is for.
