How to Write AI Video Prompts That Actually Work (With Examples)

Strong AI video prompts layer scene, camera, lighting, and one clear motion instead of piling on adjectives. This guide gives a repeatable structure and copy-ready examples that reduce artifacts and re-rolls.

Updated 2026-06-23

A reliable AI video prompt names four things in order: the scene, the camera movement, the lighting, and one primary motion, because vague or overloaded prompts force the model to guess and produce warping. The fastest improvement most people can make is to stop stacking adjectives and start describing how the virtual camera moves and what single thing changes in the frame. Specificity in those areas is what separates a usable clip from a glitchy re-roll.

Use a four-layer structure

Build every prompt from four layers: scene setup (who, where, what), camera descriptor (the move and its speed), lighting cue (direction, quality, temperature), and motion direction (what moves and how fast). Writing them in this order gives the model a clear hierarchy instead of a word soup. For example: 'A barista in a sunlit cafe; slow dolly-in from medium to close framing; warm golden-hour light from the window; steam rising gently from the cup.'

Be specific about the camera

Replace 'cinematic' with the actual move. Terms like slow dolly push, smooth orbit, steady tracking shot, crane up, and pull-back reveal directly control the virtual camera, and the more precise you are the more accurately the model executes. 'Camera orbits the watch 360 degrees and ends on a slow zoom into the dial' will out-perform 'make it look cool' every time. Motion should also be motivated: a move that reveals nothing new just wastes the shot.

Anchor start and end positions

Ambiguous motion is a top cause of artifacts. Instead of 'orbit the subject', write 'camera begins front-facing at eye level and ends at a rear three-quarter angle'. Telling the model where the move starts and stops removes guesswork and keeps geometry stable. The same applies to subject motion: 'walks from the doorway to the center of the room' beats a generic 'walking'.

Limit to one primary motion

Distortion usually comes from requesting too many simultaneous movements, overly dramatic speed, or contradictory instructions. Pick one dominant motion per clip and let everything else stay subtle. If you need both a camera move and strong subject action, slow them both down or split them into two chained clips. Simpler prompts almost always yield cleaner, more usable output.

Control speed and physics

When motion looks frantic or melts, add 'slow', 'gradual', or 'subtle' and check that your directions are physically possible. A car cannot drive and reverse in the same beat, and water cannot flow uphill, yet contradictory prompts ask exactly this. Keeping the action plausible and unhurried gives the model room to render coherent frames rather than scrambling to satisfy impossible demands.

Iterate with small edits

Treat the first output as a draft. Change one variable at a time, such as swapping the lighting cue or slowing the camera, so you can tell what actually improved the result. Keep a short library of prompt fragments that worked for camera moves and lighting setups, then reuse them. This disciplined iteration turns prompting from guesswork into a repeatable craft.

What changed in June 2026

Prompt-level control jumped across every major model in the first half of 2026, which changes how you should write prompts. Kling 3.0 (Video 3.0 Omni) launched on 5 February 2026 with a multi-shot storyboard mode: you now specify duration, shot size, perspective, narrative content and camera movement for each shot directly in the prompt, instead of stitching single clips together (PRNewswire). Google published its official Ultimate prompting guide for Veo 3.1 on the Google Cloud Blog, where Veo 3.1 follows complex multi-part prompts and generates synchronized dialogue and sound effects that you describe in the prompt itself. Runway Gen-4.5 added region-level Motion Brush control, so you paint regions on the input image and assign each an independent motion vector, letting one prompt move two subjects in opposite directions. Meanwhile OpenAI retired the Sora web and app on 26 April 2026 (the API ends 24 September 2026, per the OpenAI Help Center), so prompt workflows are migrating to Veo, Kling and Runway.

Model	New prompt control (2026)	How to invoke it in your prompt	Source date
Kling 3.0 (Omni)	Multi-shot storyboard	Label each shot with duration, shot size, perspective and camera move	Launched 5 Feb 2026
Veo 3.1	Multi-part prompt + synced audio	Describe subject, action, camera, and the exact dialogue / SFX you want	Google prompting guide, 2026
Runway Gen-4.5	Region motion vectors (Motion Brush)	Paint regions on the input image, set independent motion per region	Gen-4.5, 2026
Sora	Retired — migrate prompts	Move scene prompts to Veo, Kling or Runway	Web/app off 26 Apr 2026

Tools mentioned

Related guides

Related reports

FAQ