How to Write Image Prompts: A Universal Structural Formula

Let's start with the conclusion: nine times out of ten, a prompt that doesn't work isn't failing because it doesn't have enough words — it's failing because the order is wrong.

After all these years shooting still life, lighting has always followed an order — set the key light first, then the fill light, and only at the end do you get to softbox angle and reflectors. Writing a prompt works the same way — pin down the subject first, then add the environment, the lighting, the style, layer by layer outward, and only at the very end come the parameters. Do it backwards, pile on a heap of words, and the model has no idea whose instructions to follow.

Layer one: subject

The subject needs to be specific down to "what it is, what it's doing, what state it's in." Don't write "a girl" — write "a girl in a beige knit sweater, her face turned slightly down in profile, holding a cup of hot coffee." The more specific the detail, the less room the model has to improvise, and the more consistent the output.

This shows up clearly in the portraits category — the same phrase like "beautiful woman portrait" gives you ten different faces across ten runs; add action and clothing detail, and the stability jumps visibly.

Layer two: environment

The environment's job is to establish where the subject is and what's around them. Don't just write "indoors" — write "a wooden table by the window, an open book resting on it, blurred greenery visible outside." Environmental detail also serves a hidden purpose — it helps the model judge the depth layering of the frame, how the foreground, midground, and background should be arranged.

Clients used to ask for "keep the background cleaner, but still give it some life to it" — which, put plainly, just means the environment layer needs to spell out what stays and what gets blurred. Writing "blurred background," "shallow depth of field" directly into the prompt is faster than fiddling with parameters over and over.

Layer three: lighting

This is the layer I care about most, and also the one most people skip. Lighting determines texture, and texture determines whether an image reads as "expensive-looking." A few common phrasings:

• Side-back lighting, tracing a rim light along the edge
• Soft overhead light, no dead shadows on the face
• Natural window light, warm color temperature
• Single high-position key light, clean shadows with no muddiness

No need to stack adjectives like "beautiful lighting" — just say directly where the light comes from, whether it's hard or soft, warm or cool. The model reads light position and quality, not empty words like "beautiful."

Layer four: style

The style layer determines the overall tone — film-like, minimalist commercial photography, Japanese clean aesthetic, cyberpunk neon, and so on. This layer goes after lighting because style often ends up affecting how lighting is expressed, and getting the order wrong makes the two clash.

For e-commerce product shots, it's better to lock the style layer directly to "studio-shot solid color background," "product centered composition" — don't mix it with portrait-style wording.

Layer five: parameters

Aspect ratio, resolution, whether a negative prompt is needed — these technical items come last. The parameter layer doesn't affect the content of the image, only the output format, so it always goes last, kept separate from the content layers.

Three ready-to-copy examples

Portrait:
"A woman in a light grey wool coat, looking out a window in profile, hands in her pockets; set at a window seat in an old-style café, wooden tabletop, street view outside blurred; side-back lighting, rim light tracing the hairline, warm color temperature; film-like texture, light grain; vertical 9:16."

E-commerce product:
"A brown glass serum bottle facing the camera, the cap catching a subtle reflection; pure white background, no dead shadows, a light grey cast shadow beneath the product; overhead light with fill from both sides, no harsh reflection hotspots; minimalist commercial photography style, high contrast; square 1:1 composition."

Wallpaper/illustration:
"A mountain town after rain, bluestone paving reflecting the sky, water droplets hanging off the eaves; dusk, distant mountains tinted orange-purple by the sunset; watercolor illustration style, loose brushwork, not photorealistic; horizontal 16:9."

Copy these three, swap in your own subject wording, and keep the structure as-is.

A couple of closing thoughts

Writing prompts is a lot like lighting a shot — the more you practice, the better your instincts get, but that instinct is built on understanding the order first. Piling on words randomly is like turning on eight lights at once without thinking about where the key light should go — sure, it's exposed, but nobody's going to say it's well lit.

Want to try this out directly? The homepage has a prompt generation entry point you can fill in following this structure. And speaking of which — right after finishing this piece, the client changed their request from "make it a bit more natural" to "make it a bit more refined" — same as with prompts, saying it plainly and specifically beats endlessly revising adjectives.