Skip to content

Seedance 2.5 prompt guide: what worked for us

Keep it short. One style line, one sentence on what is happening, the beats with their seconds and the spoken lines inside them, then one line per picture and voice saying what it is. Our 1,000-character prompts beat our 5,000-character ones in blind tests.

29 October 2026 4 min read

We shoot every ShowMaker episode clip by clip, and Seedance 2.5 is one of the three video models you can pick. This is what we learned writing prompts for it, including the parts that cost us money.

The prompt that works

This is a real clip from one of our test shows: Isaac shows Dana the new smart fridge, and it has ordered forty jars of pickles. Seedance got this prompt, a picture of the kitchen, each character's sheet, a picture of the pickle jar and both voices.

Prompt
Rendered in 3D stylized CGI feature-film animation, exaggerated comedic expressions, soft volumetric lighting, smooth stylized character models, subsurface scattering, cinematic 3D render.

A sunny morning in Isaac and Dana's kitchen. Isaac proudly shows off the new smart fridge to Dana, who holds a mug of coffee.

0-6 seconds: Isaac pats the fridge door, proud: "It orders groceries. By itself." Dana, curious: "Ooh. What'd it get?"
6-9 seconds: Dana taps the fridge's touchscreen. It lights up and reads "PICKLES x40".
9-12 seconds: Dana freezes, her mug halfway to her lips, stunned: "...Forty?" Isaac folds his arms, smug: "It learns our habits."
12-16 seconds: Dana sets her mug on the island and yanks the fridge door open. Every shelf is packed with identical pickle jars. Isaac's grin stiffens as he leans in to look.

@Image1 is the kitchen. @Image2 is Isaac. @Image3 is Dana. @Image4 is the pickle jar.
@Audio1 is Isaac's voice and @Audio2 is Dana's voice: use them only for how each one sounds.

Short beat long

Before this prompt, ours were five times longer. Every beat said where each person stood, where they looked, what the camera did, and how everything ended. Each line was there because some clip had once gone wrong without it.

We shot the same clip both ways, with the same pictures and voices, and judged them without knowing which was which. In two blind tests on Seedance the short prompt won both times.

Long prompt: 4,961 characters.
Short prompt: 1,008 characters.

Pitfall

Every rule you add is one more thing the model can take too literally. Our long prompt said Dana "turns her face to him in profile". People started looking sideways instead of at each other. When one clip goes wrong, fix that clip; don't add a rule to all of them.

The rules we kept

  • Give every file one job, in one line. "@Image1 is the kitchen." "@Image2 is Isaac." The model needs to know which picture is a place, which is a person and which is a thing.
  • Say what a voice sample is for. "Use them only for how each one sounds." A voice sample is a recording of someone talking. The line tells the model to take the voice and nothing else from it.
  • Put time on the beats. Whole seconds, end to end, no gaps: "0-6 seconds", "6-9 seconds". Seedance 2.5 was built to follow timed beats, and it does.
  • Put spoken lines inside the beat where they are said. Name the speaker, then the line in quotes. Don't repeat the words at the end of the prompt; that invites subtitles.
  • Say what you want, not what to avoid. Seedance's own guide asks for positive descriptions. The two negatives it does follow: "Do not add subtitles" and "No BGM" (no background music).
  • One music decision per clip. Either describe the music once, or say there is none. On one early episode, 4 of 10 clip prompts said "no background music" and also asked for a "music sting" in the sound notes. Give music one place in the prompt.

Joining clips

A show is many clips in a row, and the joins are where it can look cheap.

  • Send the end of the previous clip. Its last frame as a picture and its last 3 seconds as a video reference. The next clip then picks up with everyone where they were.
  • End each clip at rest. If a clip ends mid-gesture, the next one tends to replay the gesture. Let the last movement finish on the last frame.
  • Don't use a start frame to join. It feels like the exact answer, but frames switch off your character sheets and voices. We wrote up why in Why we don't use start and end frames.

Refusals that cost us time

  • Realistic people. On one platform Seedance refused our own photoreal character sheet as "may contain a real person". On another, the same sheet went through. If your characters are photoreal, test them before you plan a whole episode.
  • A previous clip plus a 9:16 shape. On one platform, with a previous clip attached, Seedance once treated the request as "extend this video" and refused the 9:16 shape. It ran when we sent no shape at all.
  • A voice sample as "copyright". Once, a clip with a voice sample was refused with "audio may be related to copyright". That refusal cost us nothing.

How ShowMaker writes it for you

ShowMaker writes a short prompt for every clip from your story, sends the right pictures and voices with it, and joins the clips. You can read and edit what it wrote before anything is shot.