How to Write AI Video Prompts

One shot, one action

Most video models make clips of 2 to 10 seconds, and Kling 4.0 reaches 30. A short clip is one shot, and a shot holds one action well. "A red fox pounces into a snowdrift" fits in eight seconds; "a fox wakes up, hunts, eats and goes back to its den" does not, and the model will rush or skip the middle. When you need a sequence, write one prompt per shot and cut them together, or use the timed beats some models accept: Google's Gemini Omni Flash guide shows timecodes such as [0-3s], and Runway suggests longer clips for prompts with several sequential actions.

Subject, action, camera, light, style

Build the prompt in that order. The subject is who or what is on screen, concrete enough to picture: "a red fox in fresh snow", not "an animal". The action is a verb you can see: pounces, turns, spills, lifts off. Then the camera, then the light, then the look. A model fills every gap you leave with something generic, so each part you name is a decision you keep.

Direct the camera with film words

Models were trained on footage described in film language, so use it. A shot size says how close: wide shot, medium shot, close-up, extreme close-up. An angle says from where: low angle, high angle, over the shoulder, aerial. A move says what the camera does: static, slow push in, pull back, pan, tilt, tracking, orbit, crane up. Pick one move per clip. Two moves in eight seconds usually read as a wobble.

A suggested layout for each model

No maker requires an order. Runway's text to video guide says structure matters far less than a clear idea. These layouts are starting points, checked against each maker's pages on October 22, 2026.

• Gemini Omni Flash, in the Gemini app, Flow and the Gemini API: a scene line, then a cinematography block with camera, lighting and style. Google asks for scene, camera movement, lighting and mood, and the model makes sound, so a line of audio, such as wind or footsteps in snow, is worth adding.
• Veo 3.1, in Flow and on Google Cloud: one shot-list sentence, camera first, then subject, action, setting and look. Its model page lists clips of 4, 6 or 8 seconds.
• Kling 4.0, in the Kling app: the subject and its movement first, then the camera, the lighting and the atmosphere. Its user guide asks for descriptive words over abstract terms and negative instructions.
• Runway Gen-4.5, in the Runway web app: camera first, then the subject, the action and the setting, the structure Runway suggests to new users. From an image, spend the words on motion, since the image already shows the scene.

Sora is gone: OpenAI closed the Sora app on April 26, 2026 and its API on September 24, 2026. The video prompt generator lays your words out in each of these four orders, with nothing added but the labels.

Length and frame shape belong in the app

Every app sets clip length and frame shape in its own settings, and the setting wins over the prompt. Naming them in the prompt still helps the model pace the motion, but set them in the app too: 9:16 for a phone feed, 16:9 for a screen, 1:1 for a square post where the app offers it. Starting from an image can change the choice: Kling 4.0 takes the image's shape, and Runway Gen-4.5 offers 16:9 only from text but six shapes from an image. Shorter clips hold an action more cleanly; use the longest setting only when the action needs it.

Change one thing at a time

Treat the first clip as a draft. If the motion is right and the light is wrong, change the light and nothing else, so you learn what each word does. The same habit works for stills: the guide to AI image prompts covers it from the image side, and a strong still is often the best first frame for an image-to-video clip.