Seedance 2.5 reads a prompt as the brief for a shot, and ByteDance's own documentation says how one should be written. The BytePlus ModelArk prompt guide asks you to treat the model as "a visual content producer" and to write structured prompts with a visual storytelling mindset. 1
The guide breaks the prompt into four parts. One summary sentence sets the scene in the shape subject + location + event + genre or style + camera movement. A plot description then splits the take into segments, by timestamp or by numbered shots, and gives each segment its visuals, camera movement, actions, dialogue and sound effects. A closing block holds whatever must stay constant, such as camera angle, environment, sound and atmosphere. Every reference asset is named by its upload order and bound to a job in that text. 1
Timestamps are the part with a stated cost. The guide asks for one-second intervals as the basic unit, and for the ranges to run on without gaps: 0-3 seconds, then 3-7, then 7-15. A range holding too little plot lets the model improvise freely. A range packed too full may come back with extra cuts or missing plot. The guide keeps timestamps off high-frequency actions. 1
References follow upload order — Image 1, Video 1, Audio 1 — and each one is bound in the words. The guide's own example reads, "Image 1 depicts the protagonist John and uses the voice timbre from Audio 1." A mapping carried inside the picture instead "can easily cause character confusion or duplication". Where a reference is already accurate, naming it beats re-describing the scene it shows. 1
Negative wording has a narrow place: subtitles and audio, as in "no subtitles" or "no BGM". The guide asks for positive descriptions everywhere else. Its worked example plans a panda rolling down a forest slope as one summary line, a timeline in two ranges, and a block of consistent detail. 1 The model that prompt is written for generates a single audio-video take of up to 30 seconds and plans it against up to 50 reference assets. 2