LarpGPT · Guides
How to structure an AI photo prompt from scratch
A repeatable order for writing image prompts, what to put in each part, the common mistakes that flatten a result, and how to iterate so that each attempt teaches you something.
Why structure beats vocabulary
People collect prompt words the way they once collected search operators, and it produces inconsistent results. What produces consistent results is an order you follow every time. When every prompt has the same shape, you can compare two attempts and know what changed, and you stop rediscovering the same lessons every session.
The order that works
Subject, then setting, then moment, then light, then composition, then lens, then colour. Seven slots, one short sentence each. Not every prompt needs all seven, but writing them in this order means the early sentences constrain the later ones in the direction a photographer would work, rather than leaving the model to reconcile contradictions.
Start with a subject you could photograph
Write the subject as a physical thing with properties: material, size, condition, orientation. A weathered brass door handle is a subject. Elegance is not. The test is whether a photographer sent to shoot it would know what to look for, and if the answer is no, the model is guessing too.
Then place it somewhere specific
The setting decides what light exists and what surrounds the subject, so a vague setting undoes a precise subject. Name the kind of room, street or landscape, and one or two things that would actually be in it. Two concrete details anchor a scene better than a paragraph of atmosphere.
Name the moment
Time of day, weather, season. This sentence is unusually efficient because each of those carries an entire lighting condition. Morning after rain gives you wet surfaces, diffused light and a cool cast in four words. Skipping it leaves the model to choose, and it tends to choose bright midday, which is the least interesting light there is.
Make the light explicit
Say where it comes from and what kind it is: a window to the left, a single overhead lamp, overcast sky, direct sun. Add whether it is hard or soft. This is the sentence that most reliably separates an image that reads as photographic from one that reads as rendered, and it is the sentence most often left out.
Frame the shot
Use the words a photographer uses on set, because they describe geometry and translate cleanly. Close-up or wide. Eye level, low angle or overhead. Subject centred or off to one side. What is in the foreground and what is behind. Ambiguity here is where a good prompt produces an oddly composed result.
Add lens behaviour if it matters
A wide focal length exaggerates depth and enlarges what is near. A long focal length compresses layers. A wide aperture blurs the background and isolates the subject. Use these when the effect matters to the image and leave them out when it does not, because a lens sentence added out of habit constrains the result for no reason.
Keep colour to two or three choices
A short palette produces a coherent image and a long one produces mud. If a colour matters, attach it to an object rather than to the whole frame. A red door works. A red-tinted scene fights with whatever you said about the light.
Say what should be visible, not what it should feel like
The most common failure in a first draft is a sentence of mood. Replace each one with the physical arrangement that would produce it. This single substitution improves more prompts than any other change, because it converts something the model must interpret into something it can construct.
Avoid stacking adjectives
Beautiful, stunning, ultra-detailed, masterpiece and their relatives mostly cancel out. They do not describe anything a camera could record, and they push the result toward whatever the training data associates with the words, which is usually a generic look. Cut them and the room they free up can hold an actual decision.
Negatives are a last resort
If something unwanted keeps appearing, the usual cause is that your prompt implies it. A prompt asking for a desk often produces a laptop because desks in photographs usually have one. Adding a negative is less reliable than describing what should be there instead, since a filled space cannot also hold the thing you did not want.
Iterate one sentence at a time
When a result is close, change exactly one sentence and regenerate. You will learn what that sentence controls, and that knowledge transfers to every future prompt. Rewriting everything at once produces a better image and no understanding, which means the next brief starts from zero again.
Keep a record of what each attempt did
Save the prompt and one line on what came back. The near-misses are more valuable than the successes, because they are what you adapt when the next request is slightly different. A library of prompts without notes is a list of strings you will not trust in six months.
Reuse the structure, replace the content
To adapt a prompt that worked, swap the subject and setting first and leave light, composition and lens alone. Those sentences are what makes an image look photographic. When you do change them, change one at a time so you can tell what happened. LarpGPT provides prompts and reference material to start from, and what you generate is yours to use and yours to disclose.
One complete attempt, then one variation
Original example: vertical portrait of an adult beside a window, shoulder crop, soft natural sidelight, plain grey wall, calm expression, no text or logo. Review the resulting image by checking the person, framing and light first.
For the second attempt, keep the wording and replace only the grey wall with a blue wall. This makes the intended change easier to assess than changing the subject, lens, lighting and setting together. The generator can still vary between attempts; a prompt is not a deterministic control.
Questions
What order should I write an image prompt in?
Subject, setting, moment, light, composition, lens, colour. One short sentence each. Not every prompt needs all seven, but a fixed order lets you compare two attempts and know what changed.
Why do words like stunning and ultra-detailed not help?
They describe nothing a camera could record, so they push the result toward whatever the training data associates with them, which is usually a generic look. The space they take is better spent on an actual decision.
Do negative prompts work?
Less reliably than describing what should be there instead. If something unwanted keeps appearing, your prompt usually implies it. A filled space cannot also hold the thing you did not want.
How should I improve a prompt that is close but wrong?
Change exactly one sentence and regenerate. You learn what that sentence controls, and the knowledge transfers. Rewriting everything at once gives a better image and no understanding.
Which part of a prompt makes an image look photographic?
The light sentence, most reliably: where the light comes from, what kind it is, and whether it is hard or soft. It is also the sentence most often left out.