The Anatomy of a Perfect AI Image Prompt

Strip any great AI image prompt down and you find the same skeleton: a subject, a style, a light source, a framing choice, and a medium. Five decisions. Make them explicitly and the model gives you what you imagined. Skip them and the model makes them for you, badly.

This post walks through each part with worked examples. If you want finished prompts to copy right now, our AI image prompt library is the shortcut; this is the page that explains why those prompts work.

Part one: the subject

One noun, one or two concrete details. "A fisherman" is a category. "An elderly fisherman with weathered hands mending a net" is a picture. The detail does not need to be elaborate; it needs to be specific enough that two people reading it would imagine roughly the same person.

A useful test: if your subject could caption a thousand different photos, it is too vague.

Part two: the style

Photorealistic, watercolor, anime, 3D render, film still, charcoal sketch. This one word moves the output more than anything else in the prompt, which is why it should never be left implicit. The model has to pick a style either way; the only question is whether you chose it.

Part three: the light

Unlit prompts get the model's default: flat, even, characterless. Name a light and the image gets a mood for free. "Golden hour", "single window light", "neon signs in the rain", "overcast and soft". Photographers spend careers on this decision; you get it in three words.

A street food vendor at his stall at night, photorealistic, lit only by the stall's warm bulbs, steam rising, close-up, shot on 35mm film.

Cover the light and the medium, and even a plain subject turns atmospheric.

Part four: the framing

Close-up, wide shot, low angle, overhead, centered symmetry. Framing is how you control what the image is about. The same subject framed close reads as a portrait; framed wide it reads as a scene. If you skip it, you get a medium shot, the least deliberate choice in photography.

Part five: the medium

"Shot on 85mm lens", "Kodak film grain", "oil on canvas", "editorial magazine photograph". The medium cue ties the other four parts together and quietly sets resolution, texture, and color behavior. One or two cues are enough. Piling on "ultra detailed 8k masterpiece best quality" is a habit from older models; current ones mostly ignore it, and it crowds out words that do work.

The formula in action

Here is the same idea at three levels of control.

A woman reading in a cafe.

Model decides everything. Coin flip.

A young woman reading a worn paperback in a cafe window seat, photorealistic, soft morning light through the glass.

Subject, style, and light covered. Recognizably your idea now.

A young woman reading a worn paperback in a cafe window seat, photorealistic, soft morning light through the glass, close-up from across the table, shallow depth of field, shot on 85mm lens.

All five parts. This one comes back the way you pictured it, run after run.

When you are editing, not generating

Editing your own photo swaps one part of the formula: the subject is fixed, so the keep line takes its place. "Keep my face exactly the same" becomes the load-bearing sentence, and the other four parts describe the edit. That craft has its own guide: how to edit photos with AI prompts.

Either way, the discipline is identical. Five decisions, made on purpose. The models are good enough now that the gap between a random result and a great one is rarely the tool. It is whether the prompt made the decisions or dodged them.