How to write AI video prompts

AI video made in Root

The six parts of a video prompt

1. Shot. Wide, medium or close-up, and where the camera is: “low wide shot from across the street”.
2. Subject. Who or what, and the details that matter: “a woman in a yellow raincoat”.
3. Action. One thing that happens, from start to end: “she runs for the bus and just misses it”. One shot, one action.
4. Camera. Still, a slow push in, a pan to follow, handheld. Say how fast it moves.
5. Light and place. Time of day, weather, and what lights the scene: “wet street at night, lit by shop windows”.
6. Sound. On a model that makes sound, write the line of dialogue in quotes and name the sounds: “rain, a bus engine pulling away”.

Example prompts you can use

• “Low wide shot from across the street: a woman in a yellow raincoat runs for the bus and just misses it. The camera holds still. Wet street at night, lit by shop windows. Rain, the bus engine pulling away.”
• “Close-up of an old man's hands tying a fishing fly at a kitchen table, morning light from the window, shallow focus. A clock ticks.”
• “Aerial shot following a red car along a coast road at sunset, the camera slowly rising to show the sea. Engine and wind.”
• “Medium shot, two friends on a park bench in autumn. One says: ‘You kept it?’ The other smiles and holds up a key. Leaves, distant traffic.”
• “Macro shot of coffee poured into a glass of ice, slow motion, dark background, one hard light from the side.”
• “Handheld shot walking behind a chef through a busy kitchen to the pass, steam and flames on both sides. Shouted orders, pans.”

What each model needs from a prompt

Tips that depend on the model

• Write the sound for every model marked “On or off” or “Always” in the table: they make it from your prompt. Veo 3.1 always makes sound, so say what should be heard.
• Kling 2.5 Turbo, Runway Gen-4.5 and Luma Ray 3.2 make no sound, so leave the sound out and spend the words on the picture.
• Some models cap the prompt: Runway Gen-4.5 takes up to 1,000 characters; Kling 3.0, 3.0 Omni, 2.6 and 2.5 Turbo take up to 2,500; Luma Ray 3.2 takes up to 6,000. Keep it tight on Runway Gen-4.5.
• With a first and a last frame (the models marked Yes in the table), describe the movement between the two images, not what is in them.
• Veo 3.1 makes an 8-second clip whenever you give it a last frame or reference images, so write an action that fits 8 seconds.
• Luma Ray 3.2 makes a 5-second clip whenever you give it a first frame or a last frame, so write an action that fits 5 seconds.
• On Seedance, choose No music to ask for dialogue and effects only, and call your references by their place in the order you wired them: @Image1, @Video1, @Audio1.
• For long takes (Seedance 2.5 makes up to 30 seconds), describe what happens in order: first this, then that.

Let Root write the first draft

• The prompt box where you make a clip has three buttons. Draft turns a few words into a full prompt for the model you picked, Improve rewrites the prompt you have, and Vary gives you two or three different takes on it.
• Prompt templates set house rules for the helper, such as always shooting on long lenses, so its prompts come out sounding like you.
• In the node graph, a Describe node turns an image into a prompt that describes it, and a Prompt Enhancer node adds detail about framing, light and mood to a prompt you wired in.
• The helper uses credits, and is free when you add your own Anthropic key.

Related guides

• How to make AI video
• How to use Seedance 2
• Text to video
• Image to video
• AI image prompts
• Every guide

AI video prompt questions

Try Root free for a month