A prompt is a production brief

Failed prompts are rarely too short. They are the wrong kind of long: a pile of wishes that never decide what occupies the frame. “Cinematic, ultra detailed, masterpiece, 8k” does not describe a photograph. It describes anxiety. A working prompt is closer to a shot list. It tells the model who stands in the frame, where the camera sits, how the light is built, and what must not be “improved.” Style is a last sentence, not the job.

The slots that actually move the image

Models do not require JSON. JSON is simply a format that makes it harder to skip a field. Whether you write prose or a spec, you are filling the same slots a DP fills before a stills day:

1. The occupant of the frame: who or what, the materials, the one detail that should catch light.
2. Geometry: shot size, camera position, what is sharp, what is allowed to fall off.
3. Light as a rig: direction, quality, color, where the shadows land.
4. The world: place, time of day, what remains readable in the background.
5. Constraints: no type, no extra product, identity locked to a reference.

Leave a slot empty and the model invents it. That invention is where the “AI look” comes from: generic faces, plastic skin, logos that almost exist, rooms with no architecture. The prompts copied most often in this library look ornate. They are not. They are slot-complete. The braces are a checklist, not a trend. A prose prompt can do the same work if you refuse to skip the slots.

A rewrite

This line is common, and it decides almost nothing:

A beautiful woman in a stylish black outfit on a European street, fashion photography, cinematic lighting, 85mm, bokeh, 8k, ultra realistic

It mentions fashion photography and 85mm, then panics. The city, the hour, the pose, knit versus cut, the light direction, and whether anyone else is in the square are all left to the model. The version below is the same picture, with the decisions already made:

Medium full-body, 85mm, eye-level. Subject sharp; background falls to creamy bokeh.

A woman in a black ribbed knit set — slim cardigan, matching straight trousers — shoulder against a black lamppost, one ankle crossed, gaze just off-camera.

Daytime European pedestrian square. Pale stone buildings readable but soft. A few distant figures. Light stone pavers.

High, slightly frontal daylight with a gentle fill from camera left. Soft shadows. No hard noon pattern across the face.

True-to-color fashion still. No text, no watermark, no extra jewelry.

The format does not matter. You can compress it further. The model is not scoring your punctuation. It is scoring whether you made decisions:

shot: medium full-body, 85mm, eye-level
subject: woman, black ribbed knit set, shoulder on lamppost, ankle crossed
place: European square, pale stone, daytime, few distant figures
light: high frontal daylight, soft fill camera-left
keep: no text, no extra jewelry

The other failure: specifying everything

Once the slots are filled, you can still ruin the prompt by locking pose, micro-expression, the exact paver under the shoe, and three conflicting style names. The model starts averaging. The frame becomes a collage of instructions. Leave one degree of freedom — usually gesture, or the exact crop — so the sampler can find a photograph instead of a memo.

Quality slogans are not degrees of freedom. “Masterpiece, best quality, 8k” is noise. If you need sharpness, say the garment is sharp and the background falls off. If you need film, name grain or stock. If you need a grade, name the palette. Do not replace a judgment with an empty word.

When it is close, edit one slot

If the frame is close, do not rewrite. Change the light, or the lens, or the background — not all three. Sets come from one brief with a single node moved, not from four separate adjective paragraphs. The prompts worth copying in this library are already slot-complete. The useful edit is almost always one field, not another pile of wishes.