How to write AI prompts for images and video
Write image and video prompts with a subject, action, setting, camera intent, and constraints. Use references and one-variable tests to reduce vague wording and rework.
A useful prompt is not a pile of adjectives. It is a short brief that can be executed, judged, and revised. It identifies the subject, the main event, the setting, the intended form, and the details that must not change.
This guide works for both image and video tasks. You do not need a universal formula. Start with the smallest complete brief, keep the input fixed, and improve one variable at a time. That makes it easier to tell whether a problem comes from the prompt, the reference material, or the chosen model.
1. Define a result you can judge
Before writing the prompt, describe the deliverable in one sentence. For example: “a square skincare product image with readable front-label text, soft side light, and a light gray background,” or “a five-second vertical shot from a portrait photo, with a slight head turn, a slow push-in, and a stable face.” The sentence already establishes the output, subject, change, and acceptance criteria.
If the goal combines wardrobe changes, a new location, large movement, lip sync, complex camera motion, and exact typography, split it into shots or stages. A prompt cannot make an overloaded task easy to evaluate. Secure one clear result first, then add difficulty deliberately.
2. Build the first brief from five parts
Use subject, action, setting, form, and constraints. The subject says what matters; the action says what happens; the setting adds place, time, light, and mood; the form specifies composition, aspect, camera, or medium; constraints protect the few details that would make the result unusable.
For example: “A white ceramic coffee cup centered on a wooden table, morning light entering from a window on the left, close product photography, shallow depth of field, keep the cup logo intact.” Every phrase has a job. Repeating fashionable synonyms does not add control, and conflicting light or camera directions make the brief harder to interpret.
3. Replace vague adjectives with observable language
Words such as “cinematic,” “atmospheric,” and “high-end” communicate taste but not enough execution detail. Translate them into visible evidence: muted colors, one side light, deeper shadows, telephoto compression, a slow push-in, or generous negative space. You can then see which instruction was missed.
Avoid an unlimited negative list as well. State the positive target first, then add only failure-critical boundaries such as “do not alter the package text,” “do not add another person,” or “no camera rotation.” If the current model exposes a dedicated negative-prompt field, follow the form shown in the workbench.
4. Control composition and light in image prompts
For images, decide the subject’s scale, viewpoint, relationship to the background, and lighting before style. A product brief can specify centered or offset placement, top-down or eye-level view, and hard or soft light. A portrait brief can define head-and-shoulders or full body, gaze direction, background distance, and clothing details that must remain.
When exact text matters, reduce scene complexity, name the wording and placement, and plan for a review or editing pass. Text rendering varies by model, language, and typeface. Important brand copy should not be treated as production-ready after one unreviewed generation.
5. Add action and camera direction for video
Separate subject motion from camera motion. “The person turns toward the window” is a subject action; “the camera slowly pushes forward” is a camera action. For the first test, keep one main action in each category instead of combining running, orbiting, scene changes, and object transformation.
Give motion a direction, speed, and range. Replace “move naturally” with “the person slowly raises their head with only a slight turn while the fabric moves gently in the breeze.” Short clips need concentrated action. Duration, first-and-last-frame controls, and reference-video support must be confirmed on the current model page and form.
6. Divide the work between references and text
A reference image already supplies appearance, composition, and some style, so the prompt does not need to narrate every visible detail. Tell the model what must stay and what should change: “Keep the bottle label and composition; replace the background with wet dark stone and add a rim light from the rear right.”
A blurred, badly cropped, tiny, or contradictory reference cannot be fully repaired by wording. Improve the media first, then check how many images, videos, or audio files the selected mode accepts. Do not assume that a capability advertised for another product version is available in the current site form.
7. Change one main variable per round
After a result arrives, classify the problem: subject, composition, action, camera, style, or reference retention. Change only the corresponding instruction. If the framing is too tight, adjust the shot size; if movement is too large, reduce the range. Do not change the model, source media, prompt, and output settings at the same time.
Record the model, mode, prompt, media version, key settings, and estimated credits for each run. If the same hard requirement still fails after two or three controlled attempts, hold the other variables constant and compare another eligible model. This turns random trial and error into a reusable process.
Prompt writing FAQ
Does a longer prompt always produce a better result?
No. Extra length helps only when it removes ambiguity. Start with the subject, action, setting, form, and essential constraints, then add details that clearly affect the output.
Can I copy the same prompt into every model?
It can be a starting point, but results will differ. Models expose different modes, media inputs, and controls, so adapt the brief to the current form and run a small comparison.
What should I change first when the result is wrong?
Classify the issue as subject, composition, action, camera, style, or reference conflict. Change only the related part so the next result tells you whether that edit worked.
Write a comparable first draft for one small task
Choose an active model, run a concise structured prompt, record what happened, and change one variable in the next round.