Seedance 2.5 Prompt Guide: Writing Prompts That Work

A Seedance 2.5 prompt is one text field holding up to 20,000 characters, and it drives the picture and the soundtrack in the same pass. Anything placed in quotation marks comes back as spoken dialogue in the finished clip, which makes a prompt a shot description and a script at the same time. This Seedance 2.5 prompt guide covers how to structure that text, what belongs in reference media instead of words, and which phrasings end in a refusal.

What the prompt field accepts

The field takes up to 20,000 characters of plain text. There is no separate slot for camera notes, dialogue or exclusions, so everything travels in that one string.

A prompt is required for text-to-video. When reference images, clips or audio are attached it becomes optional, and its job shifts from describing the entire scene to directing what the references should do. A request carrying 30 reference images and three sentences of direction is a normal shape.

Four settings sit outside the prompt and cannot be set by writing about them: duration, resolution, aspect ratio and whether audio is generated. Asking for "a vertical video" in prose does not set the aspect ratio to 9:16. Set the field.

How to prompt Seedance 2.5: structure the shot

A description that can be checked against the output is worth more than one that reads well. Work through subject, what the subject does, where it happens, the light, and how the camera behaves. Every one of those is verifiable in the result, which makes a second run purposeful instead of a reroll.

A single generation runs 4 to 30 seconds and can hold several shots, so longer prompts are best written as beats. Give each beat its own sentence and say where the cut falls. Thirty seconds carries a handful of shots comfortably; twenty beats would leave each one under two seconds.

Keep the action count honest against the duration you set. Five seconds holds one action. A character who walks in, sits, pours a drink and turns to the window is asking for twenty seconds or more, and at five it arrives compressed into nonsense.

Quoted lines become speech

Audio generates alongside the picture in the same pass, covering speech, effects and music, as a mono track. Quotation marks are the control surface for dialogue.

Write each line exactly as it should be heard. Say it out loud, time it, and compare that against the duration field, because a forty-word speech does not fit in a five-second clip. When more than one character speaks, attribute the lines in the surrounding sentence so the pairing is unambiguous.

Sound that is not dialogue gets described rather than quoted. Naming the ambient bed and one or two specific effects gives the audio pass something concrete to work from.

For a clip with no dialogue, keep quotation marks out of the prompt entirely. To drop sound altogether, turn generate_audio off; it defaults to on.

Move specifics into reference media

Wardrobe, a face, a particular room, a logo shape, handwriting: those cost paragraphs to describe and still arrive wrong. One request accepts up to 30 reference images, 10 reference clips of 2 to 30 seconds each with 30 seconds combined, and 10 audio tracks.

Reference images and audio add nothing to the bill. Reference video does, and it changes how billable seconds are counted, so reach for images first.

All media is supplied as public HTTP(S) URLs. Data URIs are rejected across the platform, so host the file and pass the link.

Once references are attached, cut the descriptive prose that duplicates them. Repeating in words what an image already shows creates two sources of truth for the same detail.

Prompts that come back refused

Safety refusals return error 80006. The error does not say which stage fired, so the trigger may be the prompt, a reference file, or the generated output itself.

Three categories are blocked inside the model and no setting on any host removes them: named real celebrities, third-party intellectual property, and illegal content. Naming a living actor or a franchise character is the most common self-inflicted refusal.

content_filter defaults to true. Setting it false routes the request to a less restrictive channel and does not disable moderation; it is a routing choice rather than an off switch. The older alias for the same field is nsfw_checker.

Failed jobs are refunded in full, refusals included, so an 80006 costs time and not money. Rewriting the specific noun that triggered it beats re-running the same text.

Draft at 480p

Per-second rates scale with resolution: $0.118589 at 480p, $0.266824 at 720p and $0.461856 at 1080p for generation without a reference clip. A five-second 480p test runs about $0.59. The same test at 1080p is about $2.31.

Prompt work is structural, and structure is legible at low resolution. Draft and redraft at 480p using the duration and aspect ratio you intend to ship, then run the finished text once at 1080p. Generation takes two to five minutes either way, and polling for the result costs nothing.

Frequently asked questions

How long can a Seedance 2.5 prompt be? Up to 20,000 characters. Prompts approaching that length are usually multi-shot scripts with dialogue rather than one dense paragraph of description.

Do I still need a prompt when using reference images? No. The prompt is optional whenever reference media is present and required for text-to-video. A short direction usually still improves the result.

How do I make a character speak a specific line? Put the line in quotation marks. Quoted text generates as speech in the same pass as the picture, so keep it short enough to fit inside the duration you asked for.

Why did my prompt fail with error 80006? 80006 is a safety refusal and it does not identify which stage fired. Check the prompt for named real people and branded characters first, then the reference files, then the scene itself. The job is refunded in full.

The short version

Everything in this Seedance 2.5 prompt guide reduces to one shape: a single field, 20,000 characters, picture and sound together. Quote the dialogue, describe the rest of the sound, write the visuals as checkable beats that fit the duration, and push anything visual and specific into reference images, which are free. Leave duration, resolution, aspect ratio and audio to their own fields. Draft at 480p, ship once at 1080p.