A compelling prompt does not need to be long. It needs to be directable. A video model has to invent what a still image can leave open: time, action, camera movement, and what happens from one beat to the next. Specific, compatible choices help it make a coherent scene instead of a pile of interesting details.
Start with the image that belongs in the finished film. Describe that moment as if you were giving a cinematographer a shot brief—not asking a search engine for a collection of keywords.
1. Begin with one legible moment
Name one subject and one visible action. “A courier pauses under a glass awning as the rain eases” gives the model a person, a place and a change over time. “A beautiful, futuristic, dramatic city with a hero, a robot, a chase, a storm, fireworks and a sunset” asks for several scenes before the first one is stable.
Keep the subject easy to identify. If the scene needs a specific face, costume or object to persist, describe its defining attributes once and reuse the same language while iterating.
2. Write in camera-ready layers
Build the prompt from the parts a small film crew needs. You can start with the subject and action, then layer in the setting, framing, light, motion and mood:
- Subject: the person, creature or object the viewer should notice first.
- Action: one readable verb or short beat—turns, drifts, opens, lifts.
- Setting: place, time and a few useful environmental details.
- Framing: a wide establishing view, medium portrait, close detail, or chosen composition.
- Camera: a slow push-in, locked-off shot, gentle pan or controlled orbit—one movement at a time.
- Light and atmosphere: soft dawn backlight, cool overcast haze, crisp studio contrast.
- Finish: tactile detail, understated film grain or a clear tonal direction, when it helps.
3. Say what moves
“Cinematic” is a mood, not a movement cue. Specify what is still and what is in motion. The character can remain near the centre while a curtain moves in the breeze; the camera can drift closer as a boat crosses the background. One primary camera motion and one or two scene motions are often enough to make a clip feel intentional.
A lone cyclist crosses a quiet coastal bridge at blue hour. Wide frame, camera tracks beside them at an unhurried pace; sea fog moves over the railing, a pale cyan lighthouse beam passes across the mist. Calm, realistic, restrained colour.
The example gives the model a subject, a route, a camera relationship and a few linked sensory details. It does not prescribe every frame. Room for interpretation is useful; conflicting directions are not.
4. Choose a lens and a distance
Framing words are more actionable than adjectives alone. A wide view establishes scale; a medium view lets us read gesture; a close-up directs attention to an expression, texture or object. Pair the frame with a camera height or lens feel only when it supports the story. A low-angle portrait and a top-down product detail create very different visual priorities.
Use a small number of aesthetic references rather than stacking eras and formats. “Softly backlit 35mm film texture” is easier to interpret than “IMAX, VHS, vintage, hyperreal, stop-motion, documentary and anime.”
5. Iterate one decision at a time
Keep the strongest version of the shot brief. If the next render has the right subject but the wrong camera, change the camera line and leave the rest intact. If the light is good but the environment is noisy, simplify the setting. Changing camera, subject, lighting, style and action all at once makes it difficult to tell which edit helped.
- Make one text-only version before adding extra references.
- Keep reusable wording for a character or visual style consistent across shots.
- Change prompt details based on what the model actually did, not what a different model might support.
- Check the selected tool's duration, aspect ratio and image-reference controls; these affect the shot as much as prose does.
6. Treat model controls as part of the prompt
Different products expose different duration, camera, image-reference and editing controls. A concise prompt cannot override a setting the model does not offer. Start with the official guide for the tool you chose, use its shortest workable preview, and read the free-tier terms before submitting personal or client material.
For a current comparison of access and limitations, see the Motion Atlas directory. Credit counts and free-model access change; the official plan page remains the source of truth.
A compact template to keep
[Subject] [does one action] in [place and time]. [Frame and camera movement]. [Lighting and atmosphere]. [Tonal or material finish].
Fill in only the parts that clarify the shot. Read the prompt aloud before generating: if the camera, action and feeling point in different directions, simplify until they agree. Then keep the detail the model got right and direct the next take.