What should a video prompt include?

References, one style sentence, an avoid list, the camera, sound, and a timeline with one action per time range. Put spoken lines in quotes.

Is a longer prompt better?

Only up to a point. About one beat per two to three seconds works. Cast caps its own prompts at about 250 words so description never crowds out the action.

Should I write the prompt as a list or a paragraph?

Either works if it has timestamps. Our showcase prompts use labelled sections and a timeline; Cast's app sends one dense paragraph. The time ranges matter more than the layout.

Does the same prompt work on every model?

The structure transfers; the details do not. Reference tags, clip lengths and filters differ by model, so check each model's limits.

How do I write a prompt for an AI video?

Write it as a short production brief, not a description: what the references are, one style sentence, an avoid list, the camera and sound, then a timeline with time ranges, one action per beat and any dialogue in quotes. Video models follow timestamps and exact quoted lines far better than a paragraph of adjectives, at least in our Seedance renders.

The six parts of a video prompt

Every video we have published uses the same skeleton. It works because each part answers a question the model would otherwise guess at.

Part	What it answers	Example line
References	Who is this, and where does the shot start?	@Image1: the opening frame. Open on exactly this composition.
Visual style	What does it look like?	One sentence. If you give a start frame, the frame carries the look.
Avoid	What must not appear?	No subtitles. No logos. No change to her face.
Cast, camera, sound	Who, from where, with what audio?	One uninterrupted take, eye level, no cuts. Sound: cutlery, owls, paper tearing.
Timeline	What happens when?	0-4s: she waves the letter at the lens and says: "Reading my hate mail. Live."
Ending	How does it finish?	End on a turn: a gag, a reveal, a change of heart.

Here is an excerpt from a real prompt, from our video Hate Mail:

One uninterrupted take in continuous time through the front camera of the phone propped still on the table, eye level, the only viewpoint, no cuts, no camera movement. Posh English drawl, dry and cutting, with matching mouth movement. Sound: busy hall chatter far behind, cutlery clinks, owls hooting and wingbeats, paper tearing.

0-4s: She waves the letter at the lens: "Reading my hate mail. Live. Because I'm that hated."
4-9s: She reads the letter in a mocking voice: "'You are a spoiled brat.'" She shrugs: "Correct. Next."

How long should it be?

One beat per two to three seconds. A 5-second clip holds two beats, 10 seconds three or four, 30 seconds about five.
Around 150 to 300 words for a 10 to 15 second clip. Cast's own prompt writer caps the paragraph it sends at 1,800 characters, about 250 words, so that long character descriptions never crowd out the action.
Cut adjectives before actions. If the prompt is too long, shorten the style sentence first. The timeline and the avoid list are what keep the take usable.

How to describe the camera

Name where the camera is and what it does, then forbid everything else. Pick one of these; mixing them gives you cuts and drift.

Phone selfie / livestream: "one continuous take from the character's own phone held at arm's length, front-camera selfie framing... no cuts".
Locked-off: "locked-off observational frame, the camera only drifts, a slow lateral pan at most". Best for dances.
Push-in: "a steady dolly push straight toward the subject, tightening from medium to close as the action lands".
Low angle: "shoot from below eye line looking up... the figure made monumental".
Chase: "aggressive tracking camera travelling with the subject at speed, whip pans and snap zooms".

"In our renders, "one uninterrupted take", "the only viewpoint" and "no cuts" together stopped the model cutting to angles it made up.

Words to avoid, and words to add

Avoid franchise names, brand names and famous people. They trigger copyright filters, often after the full render. Describe things generically.
Avoid "phone" or "livestream" without an avoid line. The model may paint a phone case or a fake streaming bar. Add: "The camera itself is never visible: no phone, phone case or hand holding a phone in frame."
Avoid vague time words like "then" and "later". Use time ranges.
Add "no slow motion" if you want real-time action. Cast's prompt writer adds it to every prompt.
Add "No subtitles", or the model may draw text on screen.

If you would rather not assemble this by hand, the free video prompt builder outputs this structure. More full prompts are in the prompt library, and the Seedance guide covers model-specific limits.

Where Cast fits

You want this structure without writing it: in Cast you type one line, and an LLM expands it into a full Seedance prompt with your character, camera and vibe.
Short vertical character clips of 5, 10 and 15 seconds, where two to five beats are enough.

Where it does not:

Hand-tuning every word of the final prompt: Cast writes the full prompt for you from your line and picks.
Long multi-shot films: Cast clips top out at 15 seconds.

Make a video · first video free