To create an AI prompt from a photo, show the photo to a vision model, ask it to describe the image as a generation prompt, then cut that description down to the few traits you actually want copied. The first half takes thirty seconds. The second half is where the quality comes from.
A vision model is any model that accepts images as input alongside text. Claude and Gemini both do it, and the upload button in a chat window is the same feature the API exposes as an image content block.
The common mistake is treating whatever comes back as the finished prompt. What comes back is a caption, meaning an inventory of everything visible. A prompt is a specification, meaning a short list of things that must be true. Paste the first into an image generator and you get a picture that agrees with your photo on every detail except the one you cared about.
What follows is what the model can actually see in your photo, and which parts of its answer are worth keeping.
What a vision model sees when you hand it a photo
The model does not see the photo you uploaded. It sees a downscaled copy of it.
Anthropic's vision documentation spells out the mechanism. Claude reads images in patches instead of pixels, where each patch is a 28 by 28 pixel block called a visual token, and the token count for an image is the width in patches multiplied by the height in patches. Anything past the model's long-edge limit is scaled down first. A 3840x2160 photo arrives at a standard-tier model as 1456x819.
Fine grain is therefore gone before the model reads your question. Film grain and the small logo on a mug are the first casualties, which matters because those are often the details a person wants reproduced.
The fix is boring and effective. Send the full frame for composition and a tight crop for texture, labelling each so you can refer back to it.
The same page notes that Claude "works best when images come before text". Put the photo in the message, then the instruction underneath. It is a one-line change that costs nothing.
Three stated limits are worth knowing before you trust the description. Claude cannot determine whether an image is AI-generated. It will not name people in photographs. It cannot generate images at all, so the prompt it writes has to be carried to a separate tool.
None of these produce an error message. The model will describe your photo with total composure and no mention of the half of it that never arrived.
The parts of an AI image prompt worth extracting
An AI image prompt is a description of a picture that does not exist yet. Pulling one out of a photo means naming the few traits that carry the look, then leaving the rest unsaid.
Six slots cover almost every photograph. Ask for them by name and the answer arrives structured instead of as a paragraph you have to unpick:
| Slot | What to ask the model for |
|---|---|
| Subject | What the frame is about, in five words. A wooden chair in an empty room. |
| Shot | Apparent focal length, camera height, and where the subject sits in the frame |
| Light | Direction, hardness and colour temperature. Low sun from camera left, long shadows. |
| Colour | The grade in a few words, plus the colour the eye lands on first |
| Medium | Photograph, illustration or render, and which era or process it imitates |
| Constraint | The one thing that has to survive. Everything else is negotiable. |
Google's guidance for its image models pushes the same way, telling you to describe a scene in rich detail because "the more specific you are, the more control you have". That line gets read as permission to write two hundred adjectives, which is the opposite of what it says. Specific and long are different things.
A pile of adjectives is a negotiation the generator has to settle on its own, and it settles it differently every time you press the button. Six clear slots leave one thing to settle. Forty competing modifiers leave forty.
The constraint row is the one people skip and the one that decides whether the output is usable. Write it as the sentence you would say to a photographer who has an hour: keep the chair centred and the room empty.
AI prompt examples: the extraction prompt, and what comes back
Here is the prompt to paste under your photo. It is deliberately unglamorous.
Describe this photograph as an image generation prompt.
Use these six slots, one line each:
subject, shot, light, colour, medium, constraint.
Under 15 words per line.
Use no adjective you cannot point at in the image.The last line does most of the work. Without it you get words like "serene" and "evocative", which describe the mood in the room and give the generator nothing to draw.
Then read the answer back against the photo, line by line. Most lines will be right. One or two will be confident inventions, and they are the reason this step exists.
Generative AI prompt examples circulate as finished strings, which hides the part that matters. The useful artefact is the pair: the photo, and the prompt that got closest to it.
There is an automated version of this. Wen and colleagues' Hard Prompts Made Easy optimises discrete text prompts with gradients, letting users "mix and match image concepts without prior knowledge on how to prompt the model". The prompts it recovers from images are made of real tokens and readable in no other sense, which is a fair reminder that a prompt is input to a model whatever it looks like to you.
One more example worth running. Ask for the prompt twice on the same photo, once for a photorealistic result and once for an illustration. The two answers diverge at the medium and colour lines and agree everywhere else, which tells you which parts of your description were about the picture.
What is a negative prompt in AI, and when it does nothing
A negative prompt lists what you want kept out of the image. Several diffusion tools expose it as its own field, and it is the standard cure for extra fingers and stray watermarks.
Two things to know before you lean on it.
Support is uneven. Google's current image generation documentation makes no mention of negative prompts, so a negative carried over from an older tool lands as plain description of the thing you wanted gone.
Negation is also a weak instrument in language models generally. Our piece on cause and effect AI prompts covers the research on negated instructions, where larger models handled them worse than smaller ones. Naming the thing you want avoided puts that thing into the description.
Photo extraction produces accidental negatives, which is the failure specific to this workflow. Ask a model to describe your photo and it will report absences with the same confidence as presences: no people in frame, no text on the wall. Those lines look like constraints and are mostly noise, since your photo has a thousand absences and the model picked two of them.
Rewrite each negative as the positive it implies. "No harsh shadows" becomes "soft even light from a large source", which gives the generator somewhere to go.
Where photo-derived prompts go wrong
Three failures come up often enough to name in advance.
Likeness is the serious one. If your photo has a recognisable person in it, a prompt derived from that photo carries their face into a picture they never agreed to. Claude will not name people in images as a matter of policy, and image generators apply their own restrictions. Neither is a substitute for the consent of the person in the frame. Check the terms of the tool you generate with, and get permission before you generate someone's likeness.
Style attribution is the second. A description that lands on a living artist's name turns a craft decision into a legal one, and many tools block those names outright. Describe the technique instead, such as the brush size and the paper texture.
Seasonal edits are the third and the most common by volume. Someone has a family photo and wants an AI Christmas photo prompt that turns it into a card. Writing a prompt from scratch is the wrong tool for that job. Google's image models take the original photo plus an instruction and change only the element you name, keeping the rest unchanged, which is exactly what you want when the faces have to survive.
The extraction route is for a new picture that shares a look. The edit route is for the same picture wearing a hat. Confuse them and you end up with a Christmas card featuring a family who are almost yours.
Where to keep an AI prompt library once the prompts work
A prompt extracted from a photo has a short memory problem. Six weeks later you have the string and no idea which photo produced it or which model it was tuned against.
An AI prompt library fixes it for the cost of a few extra fields. Store each prompt next to:
- The source photo, or a link to it. The prompt is meaningless without its target.
- The model and version behind the result you kept, with the date.
- The settings that live outside the prompt text, such as aspect ratio and seed.
- One line saying what you were trying to reproduce.
That last field is the one everyone skips and the only one that cannot be reconstructed later. A prompt records your answer. It never records the question.
How to organize AI prompts has a boring correct answer. Sort them by the outcome you wanted, since the generator will be replaced long before the outcome changes. A folder named after last year's image model ages badly.
Underneath, this is an evaluation problem wearing a creative hat. Keep five source photos as fixed cases and re-run the library against them whenever a model updates. The scoring loop is the one described in our post on practices for training AI models with prompts, and it turns a model upgrade into an afternoon.
AI Prompt Engineering covers image and video prompting alongside the systems side of keeping prompts alive across model versions. Claude AI for Beginners is the gentler entry point if handing an image to a model is new territory.
The habit worth taking from all of this: once the model has described your photo, delete two thirds of it. Whatever survives the cut is the prompt that knows what the picture is for.