Why the prompt matters at all
A video model is neither a director nor an artist. It is a statistical machine: it turns your text into a shot based on what it most often saw in its training data. Give it a vague description and it fills the gaps with the average. Give it a precise one and it does what you want.
So prompting is not creativity but engineering. There is a structure that works and there are words that break everything. Let us go through both.
The formula: six blocks
Any working video prompt can be broken into six parts. You do not have to write them in exactly this order, but all six should be there.
1. Subject. Who or what is in frame. Specifically: not «a man» but «an elderly fisherman in a yellow raincoat».
2. Action. What the subject does. One action. «Slowly turns his head to the camera», not «stands up, walks and looks back».
3. Environment. Where it happens. «On a wooden pier, fog all around, boats in the distance».
4. Camera. What the camera does. One move. «Slow push-in».
5. Lighting. The character of the light. «Cold dawn light, soft shadows».
6. Style. The overall aesthetic. «Documentary footage, film grain, muted colours».
The assembled prompt: «An elderly fisherman in a yellow raincoat slowly turns his head to the camera. A wooden pier, fog, boats in the distance. Slow camera push-in. Cold dawn light, soft shadows. Documentary footage, film grain, muted colours».
See the difference from «fisherman by the sea, beautiful»? So does the model.
The main rule: one action per scene
This deserves its own section because this is where people go wrong most often.
The model generates a short segment, from 5 to 15 seconds. In that time it can properly handle one action and one camera move. Try to fit more in and you get a mess: the subject will stand and sit at the same time, the hands will live their own life, the camera will jerk in different directions.
If the story needs the hero to walk in, sit down and open a laptop, that is three scenes you will stitch together later. Each is described separately in the prompt. More tedious to write, but the result works.
A vocabulary of camera moves
Models learned these terms from millions of real videos, so they understand them more precisely than everyday descriptions.
Dolly in. The camera approaches the subject. Creates tension, involvement.
Dolly out. The camera moves away. Reveals context, a sense of an ending.
Truck / track. The camera moves parallel to the subject. Good for following a walking person.
Pan. The camera stays in place and turns. Surveying a space.
Tilt. The same, but vertically. Bottom up for grandeur, top down for oppression.
Orbit / arc. The camera moves around the subject in an arc. Impressive, but expensive in regenerations: the model gets it wrong more often.
Crane up / crane down. The camera rises or lowers. A crane, a drone.
Handheld. A slight shake. Paradoxically, it is exactly what adds realism: a perfectly stable picture reads as artificial.
If any of these terms is unfamiliar, the rule stays the same: one move per scene. «A slow push-in with slight camera shake» is still fine. «An orbit with a simultaneous rise and push-in» is asking for a reject.
A vocabulary of lighting
Light decides how a shot feels emotionally. Concrete wording instead of «beautiful light»:
Soft diffused light. An overcast day, a softbox. Even, no hard shadows. For portraits and calm scenes.
Hard directional light. Midday, a spotlight. Hard shadows, contrast. Drama, tension.
Backlight. The source behind the subject. A silhouette, a glow along the outline. Atmosphere, mystery.
Golden hour. Low warm sun. Beautiful, but worn into a cliché: use it deliberately, not out of habit.
Blue hour. Twilight after sunset. A cold palette, melancholy.
Neon. A city at night, signs, reflections on wet asphalt. Cyberpunk, the modern city.
Practical light. The source is visible in frame: a lamp, a window, a screen. Adds credibility.
Words to throw out
There is a set of filler words that migrate from prompt to prompt and mean one thing to the model: «make it averaged and glossy».
Cinematic. Describes nothing. Films come in every kind.
Atmospheric. Which atmosphere? The model will choose for you, and mediocrely.
Beautiful, stunning, masterpiece. Empty ratings. They give exactly the faceless prettiness by which AI video is recognised instantly.
High quality, 8K, ultra detailed. Incantations that barely affect the content of the shot.
Professional footage. A professional will shoot both a horror film and a shampoo ad.
Remove them. In place of each, write something concrete: which lens, where the light comes from, what imperfection is in the frame.
The secret: ask for imperfection
Real video is imperfect, and our eye knows it. The camera shakes a little, a flare hits the lens, there is grain, the focus misses for a split second.
By default a model makes things sterile. It is sterility that gives generation away. So add «defects» to the prompt:
- slight camera shake, handheld shooting,
- film grain, light noise,
- lens flares, reflections,
- shallow depth of field, soft focus,
- dust in a beam of light.
This one trick makes a clip more convincing than all the «8K» put together.
English or your own language
The honest answer: models understand English more precisely. They were trained mostly on English descriptions, so an English prompt produces fewer errors, especially on complex scenes.
That does not mean you have to agonise over translation. A working compromise: think the scene through in your own language and write the technical terms in English. Names of camera moves (dolly in, orbit), shot types (close-up, wide shot) and lighting (backlight, golden hour) are read more reliably in the original.
An image-to-video prompt is different
If you are not generating from scratch but animating a finished picture, the rules change.
The model already sees the image. So describing who is in it and how they look is unnecessary and even harmful: the text description starts to conflict with the picture, and the face drifts.
What remains in an animation prompt is only motion: what moves in frame (hair, steam, leaves, blinking) and what the camera does. And the rule of less motion is stricter here: ask for the minimum, or the anatomy falls apart. More on this in the article on how to animate a photo with AI.
Ready templates
Skeletons to plug your own details into.
Atmospheric portrait:
[who], medium close-up, slowly [one action]. [Environment]. Slow dolly in, handheld, slight shake. Backlight, soft shadows. Film grain, muted colours.
Product ad:
[product] on [surface], slow orbit around the object. Studio light from above, soft shadow under the object. Clean background, high material detail.
Cinematic landscape:
[location], wide shot, crane up. [What moves: clouds, water, grass]. Golden hour, long shadows. Atmospheric haze.
Action, dynamics:
[subject] [one sharp action], handheld, fast truck following the motion. Hard backlight, motion blur. Grain, high contrast.
Animating a photo (image-to-video):
[Motion only: the subject slowly blinks, a slight turn of the head, wind moves the hair]. Very slow dolly in. [Nothing about appearance.]
How to practise without going broke
Prompting is a skill and it is built by practice. But every generation costs money, so practise smart.
Debug the prompt on a cheap model at low resolution. At this stage you care not about picture quality but about understanding: did the model read the scene right, does the camera go where it should. That is perfectly visible at 480p too, where a generation costs several times less.
Once the prompt is debugged, move it to an expensive model and render the final once. That way one good clip costs several times less than hitting the flagship from the first attempt.
The detailed calculation of what a minute costs on each model is in the article on how much a minute of AI video costs. And if one hero repeats in your clip, be sure to read about character consistency: without references a prompt will not hold the face.
In short
A video prompt is a six-block structure: subject, action, environment, camera, lighting, style. The main rule: one action and one camera move per scene. Throw out empty words like «cinematic», add specifics and imperfection. Write technical terms in English, models understand them more precisely. And practise on cheap models, not on the flagship.
