Visual prompting means three things that get mixed together. Prompting a model to make an image. Prompting a model to make a video. And prompting a model that can see, by giving it an image and asking about it. The models differ, the prompts differ, and the habits from text prompting transfer only partly. This guide takes the three in turn, with the structure that works for each and the mistakes that show up most.
Part 1: prompting for images
An image prompt is a description, and image models read descriptions in a particular way: the first words carry the most weight, nouns carry more than adjectives, and anything you do not specify is filled from the model's average. The order that works across Midjourney, Nano Banana, DALL-E, Flux and Ideogram is subject, setting, light, camera, style. Keep it to two or three sentences; a paragraph gets its later details dropped.
A woman in a yellow raincoat (subject) standing on a wet Tokyo street at night (setting), lit by neon signs reflecting in the puddles (light), 35mm lens, eye level, shallow depth of field (camera), photorealistic, candid (style).
Three habits turn a loose description into a usable one. Replace adjectives with nouns: "beautiful lake" becomes "still lake at 6am, mist on the water, one wooden boat." Name the light, because light is what makes an image read as a photograph or a render. Name the camera even for illustrations, because "eye level, 35mm" controls the framing whatever the style. Then say what to leave out in a short plain sentence at the end: "no text, no other people."
For a series, write the setting, light, camera and style once and change only the subject. The series matches because everything but the subject is identical. For a real product or person, attach a reference and say what to keep: "keep the face, hair and label exactly as in the reference; change only the setting."
Part 2: prompting for video
Video prompts add motion and time, and the models (Sora, Veo, Runway, Kling and the rest) are strict about both. One action per clip: a second action becomes a cut, and the cut is never where you want it. State the camera move separately from the subject's action, state the duration, and name the sound if the model generates it, or it will invent one.
A cyclist in a red jacket rides through a wet city street at dusk (subject and action), camera tracking alongside at the same speed (camera move), neon reflections in the puddles, 35mm, shallow depth of field (look), sound: tyres on wet asphalt, distant traffic, no music (audio), 8 seconds (duration).
Keep the prompt under about 80 words. Keep one constant visual detail on the subject (the red jacket) so the same character can appear in the next clip. Some video models read a structured prompt more reliably than prose; a JSON block with keys for shot, subject, action, setting, light, audio and duration works in Veo and is readable by the others. For a sequence, write one prompt per shot with the subject line identical across them, and edit the clips together afterwards; no current model builds a multi-shot scene from one prompt well.
Part 3: prompting with an image
The third kind is the one people forget is visual prompting: giving ChatGPT, Claude or Gemini an image and asking about it. The model reads the image and your text together, and the text decides what it looks for. "What is this?" gets a caption. A specific question gets analysis.
The structure is: say what the image is, say what you want from it, say the format. The model sees the image but does not know why you are showing it, and "this is a screenshot of my dashboard" changes what it attends to.
This is a screenshot of our checkout page on mobile. List every element a first-time buyer would have to find before they could pay, in the order they appear, and mark the three that are hardest to find. Then suggest one change for each. Table format.
This is a photo of a whiteboard from our planning meeting. Transcribe everything written on it, keep the groupings as they are drawn, and mark anything you cannot read with [unclear].
This is a chart from a report I do not have the data for. Describe what it shows in three sentences, estimate the values at the labelled points, and say what the chart does not let you conclude.
Uses that work well: transcribing handwriting and whiteboards, reading a chart or a table from a screenshot, describing a UI for an accessibility review, extracting text from a photo of a document, comparing two screenshots for differences, identifying what is in a photo for a listing description, and getting a prompt back from an image you like ("write the prompt that would produce this image in Midjourney"). Uses that need care: anything where a wrong reading has a cost, such as medical images or legal documents, where the model's confidence does not track its accuracy. Ask it to quote what it sees and mark uncertainty; it will.
The mistakes that cross all three
Vague adjectives. "Cinematic", "beautiful", "professional" mean the model's average. Replace with the noun you mean.
Too many instructions. Image and video models drop the later ones; vision models answer the last question. One prompt, one job.
Text in images. Models misspell. Leave words out of the image and add them in a design tool, or use Ideogram, which handles short text best, and keep it to a few words.
Asking for the model's opinion instead of its observation. "Is this design good?" gets a polite answer. "List what a user would struggle with" gets a useful one.
Where to go next
The art and design prompts are the image half of this guide in tested, model-specific form, grouped by what you are making. For video, the Veo mastery guide covers the JSON structure and the one-action rule in depth, and the Sora 2 prompts show the prose version. For prompting with an image, the image-to-prompt guide covers the reverse direction, getting a prompt from a picture, and the data visualization prompts include reading charts back into numbers.
