Seedance Prompt

seedance-prompt

Before producing prompt text, a prompt-ready block, a rewrite, an example, or a compiled clip, load the Director's Read, classify the brief, and complete its canonical narrative or non-narrative record. Translate that record into visible or audible carriers and keep its internal labels out of final generation prose.

Build production-ready Seedance prompts from clear concepts or supplied reference assets. Treat the prompt as a short shooting brief: it must say what changes on screen, what the camera does, what the light and sound contribute, and what must stay stable. Keep final prompts under the platform prompt budget and remove filler before delivery.

Load the Director's Read before any drafting or compression, quick-ref for the checklist, reference-workflow for multimodal references, i2v-guide for image-to-video, first-last-frame-guide for first/last-frame work, examples-by-mode when examples are useful, shot-list-continuity for multi-shot professional plans, multishot-grammar for shot order, the continuous-versus-storyboard shape rule, the load-per-beat ladder, and cut placement inside one generation, and multilingual-community-examples for Chinese/Russian/Japanese/Korean/Spanish or mixed-language prompts. When sequence state is present, load prompt-compiler and compile only the current clip contract.

Intent

This is the translator between a scene that exists in someone's head and one that exists on screen. The user has already imagined it; the job is to lose as little as possible in transit. Success is a first generation close enough that they can react instead of explain. Each revision inherits everything the story has already decided and changes only what the reaction asked for - a draft is a conversation, not a restart.

Director Formula

Before filling slots, run the Director's Read gate. A brief with a dramatic story turn must have all ten canonical narrative fields completed. Observation/performance without that turn and utility briefs use the two-line non-narrative intent and refusal, with no fabricated psychology. Preserve expressive movement, rhythm, or delight when requested. For narrative work, name a single intention from the read and let it choose camera, lighting, blocking, performance, and sound together. Load directing-engine when the setup needs its deeper instrument and voice logic. The formula below is the container for a coherent setup, not a checklist of independent decorations; if a project voice is already set, keep this shot inside it.

The Director's Read is internal planning, not prompt copy. For narrative work, compile its turn, visible suppressed behavior, and non-transferable detail into visible or audible carriers. For observation/performance or utility in the non-narrative lane, carry the requested action, rhythm, information, or sensory change without inventing those dramatic fields. Do not paste labels such as POV, power shift, hidden want, or subtext into the generation prompt, and do not ship an abstract feeling where blocking, gesture, prop use, camera endpoint, motivated light, dialogue contradiction, silence, or a sound cue can carry it.

Use Subject + Action + Scene + Camera + Lighting/Style + Audio + Constraints. Put the subject and primary action first because early clauses set the shot hierarchy. Do not force every slot if a reference asset already shows the information; for I2V, describe only the motion, camera, timing, transformation, audio, and preservation constraints that the still image cannot show.

Mode Gate

Choose the mode before drafting. T2V needs subject, action, scene, camera, light, style, and constraints because nothing is visible yet. I2V starts from @Image1 and adds only motion, time, camera, lighting transition, audio, and preservation. V2V should map @Video1 to source clip, camera move, action rhythm, blocking, edit target, or extension anchor rather than accidentally transferring identity. R2V must list every reference role and state what must not transfer. FLF2V uses @Image1 as first frame and @Image2 as last frame, then describes only the continuous transition.

Time Structure

Seedance 2.0 keys on shot order, not seconds (official guidance recorded 2026-09-26; sources and the full rule in multishot-grammar). Classify the shape first. One scene with one continuous action or state change is continuous and compiles as one paragraph with no shot labels, whatever its duration. Several events, a location change, a reveal that needs a cut, or a comparison is a storyboard and compiles as numbered shots in event order, each block ordered camera move or cut, action and expression, space change, audio, with the cut written in words. Never write absolute seconds such as 0–3 s inside a shot block on 2.0; duration is the surface parameter, and a felt length is written as behaviour.

For a storyboard clip, score the load (a camera move, each spoken line, extra people, contact that must land, a location change) and place the requested shot count on the ladder: Safe, Stretch, or Ambitious. When the shot count is open, present up to three rungs with the trade-off in one line each, recommend one (Safe on a first attempt or a last credit, Stretch when the user has room to iterate), say once that the thresholds are this skill's heuristics and that no official shot ceiling exists, and write one finished prompt at the recommended or chosen rung; other rungs are written on request, and "choose for me" means draft the recommendation. When the user has fixed the count, write it and state its rung in one line. Ambitious is never called broken; two shorter generations are offered as the better spend. A request for timestamps on the newer model line stays inside the boundary in api-status: keep the craft, withhold the numbers.

Sequence Boundary

The generic prompt skill must not independently invent continuation state. If the user asks to continue, extend, make part two, or use a previous clip, route to seedance-continuation unless the accepted clip/final frame and observed end state are already present in the sequence state.

For sequence prompts, preserve project_id, clip_id, parent_clip_id, continuity locks, exact reference tags, the actual opening state, completed beat exclusions, and reserved future beats. The final prompt remains natural language and covers only the current clip.