How to write prompts for AI video: shot, subject, motion, camera and style
A five-part structure for AI video prompts, with before-and-after examples you can paste into Text to Video, plus notes on length, audio and the Camera presets.
A video prompt is closer to a shot description than to an image caption. An image is a single moment. A video has a beginning, a middle and an end, and the model has to decide what moves, how fast and from where it is filmed. If you leave those choices open, the model makes them for you. This guide gives you a structure for making them yourself.
The five parts of a video prompt
Write one or two sentences that cover these parts, in roughly this order:
- Shot. The framing: close-up, medium shot, wide shot, overhead, low angle.
- Subject. Who or what is in the frame, with two or three concrete details.
- Motion. What changes over the clip. This is the part people forget.
- Camera. Whether the camera is still, pushing in, panning, tracking or handheld.
- Style and light. Time of day, light quality, colour mood, film look.
Here is the difference it makes.
A weak prompt:
A woman in a city at nightA prompt with all five parts:
Medium shot of a woman in a yellow raincoat walking along a wet city street at night, neon signs reflected in the puddles, she glances over her shoulder and keeps walking, slow tracking shot from the front, cinematic, shallow depth of field, cool blue and pink lightThe second prompt tells the model what the frame looks like, what the subject does and how the camera behaves. There is much less left to chance.
Describe motion, not just appearance
Appearance is easy to write and usually not the problem. Motion is where video prompts succeed or fail. Use verbs, and say how fast:
- "slowly turns her head toward the window"
- "steam rises and curls away from the cup"
- "waves roll in and break against the rocks"
- "the crowd walks past in the background, out of focus"
Limit the amount of action. A 5 second clip can hold one clear action and perhaps a small second one. "A man runs, jumps over a fence, catches a ball, and celebrates" is four shots of content in one clip, and the result will usually be rushed or confused. Split it into separate clips.
Say what the camera does
Cameras are a strong lever because they change the feel of the shot without changing the subject. Useful phrases:
- Static camera, locked off. Calm, documentary, good for product shots.
- Slow push-in or slow pull-back. Adds emphasis or reveals context.
- Pan left or right, tilt up or down. Follows or discovers.
- Tracking shot. The camera moves alongside the subject.
- Orbit around the subject. Good for products and sculptures.
- Handheld. Adds energy and realism.
- Drone fly-through or crane shot. Larger scale moves.
In the Video tools, the Camera button next to the prompt opens a Camera movement panel with presets such as Pan Left, Zoom In, Tracking Shot, Orbit 360 and Crane Shot. The panel notes that these presets work best with Kling and Seedance models. A chosen preset is added to your prompt as a short "camera:" phrase, and you can always write the movement in your own words instead.
Do not ask for several camera moves at once. "Pan left while zooming in and orbiting" tends to produce a muddle. Pick one.
Add style and light in a few words
Style words work best when they are specific and few. Compare "beautiful, amazing, stunning" with "golden hour backlight, soft haze, 35mm film look". The second set gives the model something to act on.
Reliable choices include time of day (dawn, noon, blue hour), light quality (soft window light, hard sunlight, neon), lens feel (wide angle, macro, shallow depth of field) and a single mood word.
Dialogue, sound and the Audio switch
Several of the video models generate sound together with the picture. In the tool, the Audio row has an on/off switch, and for some models turning it off lowers the price. If you plan to add your own voice and music, switch it off. If you want ambient sound, describe it in the prompt: "distant traffic, a dog barks once, wind in the trees".
For spoken lines, put the words in quotation marks and say who speaks:
Close-up of a chef tasting soup from a wooden spoon in a busy kitchen, he nods and says "Needs a little more salt", steam rising, handheld camera, warm tungsten lightSpeech in generated video is less predictable than a dedicated voice tool. For a script you must control, make the video silent and add a recorded voice-over. The voice-over and music guide shows how.
Length, resolution and model
- Duration. The models available in FlexPic accept different ranges: Seedance 4 to 15 seconds, Kling 3.0 3 to 15 seconds, Veo 3.1 4, 6 or 8 seconds, Wan 2.7 2 to 15 seconds. Longer clips cost proportionally more and give the model more time to drift away from your idea.
- Resolution. Draft at 480p or 720p. A 5 second Seedance 2.0 clip costs about 444 credits at 480p and 2,363 at 1080p.
- Aspect ratio. Choose it before you generate, because it changes the composition. See aspect ratios for social platforms.
The Auto Polish switch under the prompt appends the words "highly detailed, cinematic lighting" to what you wrote. It is a convenience for very short prompts. Turn it off when you have already described the light and style yourself, so the extra words do not fight your own.
A working routine
- Write the five-part prompt. Keep it under about 60 words.
- Run it once at the lowest resolution and a short duration.
- Watch it twice. Note one thing to fix.
- Change only that one thing and run again.
- When the motion and framing are right, rerun at final resolution.
If a result keeps missing the same way, the troubleshooting guide lists the common failures and what to change. If you already have a still image you love, image to video gives you more control than text alone.
Try it
Open Text to Video, paste the tracking shot prompt from above, set a 5 second clip at a low resolution and compare two models on the same text. The difference between models is often easier to see than to read about.