Prompting

How to give a model direction in words, for pictures and for video.

A prompt tells a model what you want in words. Words work well for the subject, the mood, the action and what the result is for. For things that are easier to show, such as a face, a product or a layout, connect a reference instead; see Working with references.

Give it structure

Say what the result is for, such as an ad, a product shot or a storyboard frame. Then describe the subject, the setting, the light, the framing, and any space that should stay empty. A few short lines are easier to work with than one dense paragraph.

Say what must stay

When you edit a picture, say what should change and what must stay the same: the face, the product's shape, the label, the layout. Then check those details in the result.

Change one thing at a time

Start simple, look at the result, then adjust one thing: warmer light, more space on the left. If something goes wrong, you'll know what caused it.

Prompting video

A video prompt describes one shot, for a reader who has seen nothing else. Write in this order: who or what is in the frame, what they do, where and in what light, how the camera sees it, and what's heard. Use the present tense, and keep to one action and one camera move; a five-second clip holds one gesture and one line. Forty to 120 words is plenty.

If a reference is connected, don't describe what it shows; say what it is and what to keep. If an image is on the Start frame port, it becomes the first frame: don't describe it, and write only what changes.

Dialogue and sound

Put spoken words in quotes after the speaker's name:

Tom says: "You came." Then Sara replies quietly: "I said I would."

Only the words inside quotes are spoken, at about two and a half words a second, so a five-second clip fits one line of about ten words. For a silent clip, write No dialogue.

A video model with audio on always makes sound. One sentence directs it:

Sound: wind in the trees, gravel underfoot. No music.

Camera and negatives

Give each shot one size (wide, medium, close-up), one angle (eye level, low, over the shoulder) and one movement (static, slow push-in, pan, tracking). Across shots, keep people on the same side of the frame, and say so in every prompt.

Only use negatives for things models add on their own, such as subtitles, on-screen text or music, on one line at the end. For everything else, say what you want instead.

A complete video prompt

A five-second clip with a photo of each person connected:

@Image1 is the man. Call him Tom. Keep his face, hair and clothes exactly as shown. @Image2 is the woman. Call her Sara. Keep her face, hair and coat exactly as shown. A park bench in late afternoon sun, trees behind. Tom sits on the left end of the bench; Sara stands on the right, just arrived. He looks up from his phone and says warmly: "You came." Then Sara replies quietly: "I said I would." Medium two-shot, eye level, static camera. Sound: light wind in the leaves, distant children. No music, no subtitles.

Further reading

The OpenAI image prompting guide has detailed examples. Most of its advice works for other models too.

Guide version 2026.10.21-3 · October 21, 2026