Guides
Writing a prompt a model can film
Video models are not image models with time added. They respond to how a shot is described — subject, camera, light, in that order — and there are a few things they are still reliably bad at.
Say what is in the frame before you say how it looks
Models weight the beginning of a prompt more heavily. Lead with the subject and the action, then the setting, then the treatment.
Weaker: Cinematic, moody, anamorphic, 35mm, a woman.
Stronger: A woman turns to look over her shoulder, in a rain-slick alley at night, anamorphic flares.
The second one has something to film. The first is a list of moods with a subject appended.
Describe the camera as a move, not as a lens
"Slow dolly in", "handheld push", "static wide", "crane up" — these are instructions a model can follow because they describe motion over time. "Shot on ARRI Alexa" describes a look, not a movement, and models mostly interpret it as "make it a bit more filmic".
One camera move per clip. Two competing moves in five seconds generally produces neither.
Light is the cheapest way to change the result
Changing nothing but the light — "overcast", "hard midday sun", "single practical lamp", "backlit at golden hour" — moves the output more than most adjectives will. It is also the fastest thing to iterate on, because it does not change the subject or the composition you already liked.
What models are still bad at
- Text in frame. Signs, labels, screens. It will usually be almost-words. Plan for it or crop around it.
- Hands doing something specific. Holding an object is fine; operating a mechanism is not.
- Counting. "Three people" often gives two or four. If the number matters, frame so it is unambiguous.
- Continuity across clips. The same character in two separate generations will not match. For a consistent presenter, use a saved presenter and lip-sync rather than describing them twice.
Iterate cheaply, then commit
Find the shot at the smallest size and shortest length the model offers — that is what the cheap tiers are for. A 360p five second test costs a fraction of a 1080p ten second one and tells you almost everything about whether the composition works.
When the prompt is right, raise the size and length and generate once. Nothing about the prompt needs to change.
When to compare instead of guess
Models disagree most on style, not on competence. If a prompt is giving you something technically fine but wrong in feel, run the same prompt through two or three models side by side rather than rewriting it — the difference between models is often larger than the difference another adjective will make.
The comparison shows the total cost before it runs, because three models means three charges.