How to write prompts to generate videos
How to describe a shot for Runway, Kling, Sora or Veo: one subject, one action, camera movement, light and mood. And how to plan multi-shot videos.
Generating video with AI is more like directing a film shot than writing text. The model needs to know what is seen, what moves and how the camera moves.
One shot, one action
Video generators produce short clips. They work best when each prompt describes a single shot with one main action. If you need a story, split it into shots and generate each one separately.
Recommended structure
- Shot type: wide shot, medium shot, close-up, aerial, low angle.
- Subject: a concrete visual description.
- Action: what it does, at what speed.
- Setting: place, time of day, weather.
- Camera movement: static, tracking, slow push-in, pan, handheld.
- Lighting and mood: soft overcast light, sunset, neon, nostalgic or tense mood.
Example
Close-up shot of an elderly baker's hands kneading dough on a floured
wooden table, slow and steady movements, early morning in a small bakery,
camera slowly pushes in, warm light from a window on the right,
flour particles floating in the air, calm and nostalgic atmosphere
Text-to-video and image-to-video
- Text-to-video: you describe the whole scene.
- Image-to-video: you start from an image (generated or your own) and the prompt focuses on motion: “the camera slowly pulls back while the wind moves her hair”.
The second option gives more control over the final look and helps keep shots consistent.
Continuity tips
- Describe the character with the same words in every shot.
- Use the same palette and lighting.
- If the tool supports reference images or saved characters, use them.
Script before prompts
For ads or social videos, write the script and storyboard first (a text assistant can help) and then turn each shot into a video prompt. The library has a prompt for an ad script and storyboard.
Keep in mind
Length, resolution and aspect ratio are set in each tool's interface, not in the prompt. Results vary a lot between attempts: generate several versions. Respect people's image rights and each platform's terms of use.
More examples in video creation prompts.