Should I use timestamps in Seedance prompts?

Yes. In our renders, beats labelled 0-4s, 4-9s and so on were followed far more reliably than the same events written as prose.

How do I add dialogue to a Seedance prompt?

Put the exact words in quotes inside the beat where they are spoken, and describe the voice once, such as "posh English drawl, with matching mouth movement".

How long should a Seedance prompt be?

Long enough to cover every beat, usually a few hundred words for 30 seconds. Length is fine; vagueness is the problem.

Do Seedance 2.0 and 2.5 need different prompts?

The same structure works for both. Size the timeline to the model: up to 15 seconds on 2.0 Fast and up to 30 on 2.5.

How do I write a good Seedance prompt?

Write it as a timeline, not a paragraph: label each reference, add one line of visual style and an avoid list, say exactly where the camera is, then list beats with timestamps (0-4s, 4-9s) and put every spoken line in quotes. In our renders Seedance followed timestamps far better than descriptive prose.

The structure that works

Every video in our prompt library uses the same five parts, in this order:

References. One line per attached file: what it is and what to keep. "@Image1: the opening frame. Open on exactly this composition." Add a line per extra face and "@Audio1" for a song.
Visual style. One sentence. If you start from a painted first frame, the frame already carries the look; do not re-describe it at length.
Avoid list. Plain negatives: no subtitles, no on-screen interface, no phone in frame, no change to her face.
Who, where, camera, voice, sound. One sentence on who the character is. One on where the camera physically is, with "one uninterrupted take" and "no cuts". A voice description. A "Sound:" line with three to five sounds.
The timeline. Beats with time ranges, one action and at most one line each, ending on a turn: a gag, a reveal or a change of heart.

This is the opening of the prompt behind Hate mail, a 30-second Seedance 2.5 video:

A spoiled, aristocratic, bossy, saucy rich girl reads her hate mail live at breakfast. 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.

Shot 1 (0-30s): [Front camera of her phone on the breakfast table, one continuous take]
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."

Timestamps, dialogue and beats

Use timestamps. One beat per 3 to 6 seconds is a good density: a 30-second take holds about five beats, a 5-second clip one or two.
Quote every line exactly. Lines in quotes come out spoken and lip-synced in the voice you describe. Keep each line short enough to say in its time range.
One action per beat. "She opens the letter, reads it, laughs, burns it and calls her friend" in four seconds gets compressed or dropped.
Give each beat a sound. It keeps a static camera from feeling static.
Narration instead of dialogue? Write "Nobody speaks; mouths closed or smiling." Otherwise the model invents lip movement under a voice-over it never heard.

What the avoid list should say

Keep it short and concrete. Our default is below. Delete "No background music" when you pass a song as a reference, because it fights the audio. Add "No logos, crests or numbers" for any uniform or kit.

Avoid: No subtitles. No background music. No streaming-interface overlays, phone filters, or masks. No phone body or hand holding a phone in frame or in reflections. No change to her face.

If the camera is a phone, Seedance sometimes paints the phone case into the foreground. Adding "The camera itself is never visible: no phone, phone case, tripod or screen in the foreground" fixed it for us.

Mistakes that waste renders

Franchise names, famous characters or props: in our renders the copyright filter blocked them after the full render. See why Seedance blocked my video.
Describing a face that reads as a famous actor.
Letting the model choose the camera: it invents cuts. Name the camera position and say it is the only viewpoint.
Writing more than the clip can hold. Seedance 2.0 Fast stops at 15 seconds per clip.

For a filled-in skeleton, try the free video prompt builder. The full set of our lessons, including chaining clips past 15 seconds, is in the Seedance field guide.

Where Cast fits

You would rather not write the timeline yourself: Cast turns one line plus a style, camera and vibe pick into a full Seedance prompt with a language model, then renders it on Seedance 2.0 Fast or Seedance 2.5.
You want to learn by example: every published showcase video has its exact prompt on its page.

Where it does not:

Hand-editing every word of the final prompt before it renders: Cast writes the prompt from your line.
Prompts that need an uploaded song, video or your own image as a reference.

Make a video · first video free