How to write prompts for AI video generators
Video models read prompts differently from image models. An image prompt describes a moment; a video prompt describes a shot. Once you write it like a shot list, results from Kling, Veo, Sora, Luma and the rest become predictable instead of random.
The structure that works
Use the same order every time: shot type, subject, action, camera movement, lighting, style, duration feel. For example: 'Medium shot of an elderly fisherman mending a net on a wooden pier, hands moving steadily, slow dolly-in, golden hour backlight, documentary realism.' Every element answers a question the model would otherwise guess.
Keep one action per clip. Models handle one continuous motion well and chains of actions badly. If you need a sequence, make three clips and cut them together in the project.
What causes flicker and morphing
Contradictions: 'static camera' plus 'sweeping aerial', or a style word that fights the subject. Vague verbs: 'doing things' gives the model nothing. Crowds and hands in close-up: still the weakest point of every model. And overlong prompts: past about sixty words, later words are mostly ignored.
Negative prompts help on models that support them ('no text, no extra limbs, no warping'), and on all models a lower motion setting trades energy for stability.
Model differences you should know
Models are not interchangeable. Some are strong at realistic human motion, some at camera work, some at stylised animation, and prices per second vary several-fold. The useful habit is to run the same prompt on two or three models and keep the best, which only works if the tool lets you switch models without re-uploading everything.
How to do it in Vyndexo
Write the shot
In Media Studio, write the prompt in the shot-list order above. Save it in the project so it can be reused.
Run it on two models
Choose a model, generate, then switch to a second model and generate again with the same prompt. Costs are shown per model before you start.
Compare in the project
Both clips land in the same project with their prompts and settings attached, so you can see what each model did with identical input.
Refine one variable
Change one element at a time (camera move, light, style) and regenerate. Failed generations are not charged.
Common questions
How long should a video prompt be?
Thirty to sixty words. Enough to specify shot, subject, action, camera and light; short enough that every word counts.
Do the same prompts work on every model?
The structure does; the exact wording may need small changes. Running the same prompt across models side by side is the fastest way to learn each model's preferences.
Can I use a reference image instead of describing everything?
Yes, and it is usually better. See the photo-to-video guide.