How to write a video prompt

A prompt is the description of a clip. It reads best like a shot description, not a title. Say what we can see, what moves and where the camera goes.

Describe the clip, not the idea

"A cool coffee ad" gives the model nothing to draw. "A cup of coffee on a wooden table at golden hour, steam rising" does.

Write what a camera would see: the place, the light, the person and what they do.

Weak
A beautiful, premium video of coffee.
Better
Live-action, a medium shot of a white cup of coffee on a wooden table at golden hour. The camera pushes in slowly toward the cup. Steam rises from the cup.

Start with the style and the distance

Open with the look: live-action, 3D animated, vintage film. Then the distance: close-up, medium shot or wide shot.

Everything else hangs on these two choices, so they go first.

Name people and places with things you can see

Skip words like "beautiful" or "premium". Use age, clothes, hair, the room and the light.

For an Indonesian scene, say so: a warung counter, a kos room with a standing fan, a woman in a hijab, an angkot at dusk. A specific place looks more real than "a street".

One action, one camera move

Give each shot one action: she lifts the bottle, the cyclist rides through the puddle. Write the camera as a sentence. Do not stack labels like "slow push-in, static tripod", because they contradict each other.

Camera sentences you can reuse:

  • The camera pushes in slowly toward the cup.
  • The camera holds a static shot.
  • The camera pans right slowly.
  • The camera shakes slightly, like a phone held at arm's length.

Say what is there, not what to avoid

Naming something you do not want can bring it into the clip. Instead of "no text on the bottle", write "the label is plain white".

Use the three-part form where a model reads sound and music

Some models, Senfers 1.0 among them, read a prompt in three parts: the scene, the sound and the music. Write the scene as above, add a line for the sounds in the room, and a line for music, or N/A for none.

Other models take a plain sentence or two.

Scene, sound, music
integrated_multimodal_description: [Shot 1] Live-action, a medium shot of a white cup of coffee on a wooden table at golden hour, a blurred window behind it. The camera pushes in with small amplitude at slow speed toward the cup. Steam rises from the cup.

overall_soundscape: Low café murmur and a spoon touching ceramic.

non_diegetic_music: N/A

Leave the three labels at the start of each part as they are. The model reads them.

Write the prompt in English

Prompts in English give more accurate results, also for an Indonesian scene. Keep the local details: the warung, the kos room, the batik sleeve.

Try short, then go longer

Test an idea at a short length first. When the action looks right, make the longer version. Each model page lists its lengths.

Write your own prompt