Seedance Sequence

seedance-sequence

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.

Use this when the user's idea is larger than one reliable generation, when connected clips are requested, or when the user says continue, extend, next part, part two, next scene, or make it longer. Plan globally, generate locally: the skill plans the whole story, but compiles only the next unresolved clip.

Load the Director's Read, sequence-project-state, continuation-handoff, prompt-compiler, surface-prompt-profiles, event-density, and continuity-qc. Load reference-transfer-contract when references are present and dense-storyboard-mode when the request contains many shots or animation panels. Load directing-engine to set one directorial voice for the whole story and plan the long-form spine so the look is authored by one hand across every clip. For a user's first multi-clip project, sequence-worked-trace walks the whole loop once - plan, deviation, reconciliation, chain cap, re-anchor, resume.

Intent

The user is trying to make a film, not a pile of prompts. This skill protects the thread of action across generations: what already happened, what is happening now, what must not happen yet, and what the accepted footage actually shows. The plan is global; the prompt is local.

Sequence Classifier

Classify as sequence_project when the story exceeds the verified active-surface duration, asks for multiple connected clips, contains several narrative beats, is a film scene, ad, campaign, music sequence, action scene, dialogue scene, or uses continue/extend/next-part language. Otherwise classify as standalone_clip and return to the concise prompt path.

For every request also classify:

• generation input mode: T2V, I2V, V2V, R2V, FLF2V, edit, native extend when verified for the active surface, or troubleshoot;
• sequence relation: standalone, sequence_first_clip, seamless_continuation, intentional_next_shot, bridge_between_known_states, repair_tail, or reanchor_after_drift;
• shot structure: compact_single_take, phased_single_take, dense_multishot, first_last_frame_transition, or video_edit_contract;
• medium grammar: live_action, 3d_animation, 2d_animation, product_or_object, or another supported medium;
• surface profile: exact reference-tag convention, verified duration range, prompt budget, supported reference roles, timeline syntax, edit/extension availability, audio behavior, and constraints.

If the surface is unknown, use a conservative generic profile. Do not invent a duration, prompt limit, reference count, or tag syntax.

Scene Architecture

Plan scenes before clips. A scene is the re-anchor unit: one location and time envelope whose clips may chain from each other's accepted footage.

• Seamless continuation is legal only inside a scene. A scene boundary is an intentional cut that opens from canonical references and resets extension_depth to 0.
• Cap consecutive output-sourced generations at the scene's max_chain_depth (default 2, hard ceiling 3). Schedule re-anchors in the plan; identity decays with chained generations, so a scheduled reset is routine and a drift repair is expensive.
• Map the arc to scenes: each scene carries one arc_position and its clips inherit it.
• Cuts are the cheapest continuity tool. The audience expects frame continuity only inside a chained shot, not across an editorial cut. A five-minute story usually resolves to several scenes of two to five clips, not one long extension chain.
• Audio: clips carry ambience, sync SFX, and on-camera dialogue; unify music and score in post because audio is not continuous across separate generations.

Build Process

1. Establish the story promise and final outcome before Clip 01.
2. Classify every clip with the Director's Read and record its explicit directors_read_lane as narrative or non_narrative; never infer psychology from story.medium. This per-clip lane lets utility inserts coexist with performed beats in one sequence. Before a clip classified in the narrative lane is compiled, complete its ten-field internal record, including one non-transferable detail and one stock-solution refusal. For observation/performance without a dramatic story turn or for utility, use the non_narrative lane and record the two-line intent and refusal; do not invent drama or suppress the requested expression. Then use directing-engine to set one directorial voice for the whole project when the sequence needs it and plan the long-form spine - how shot scale, camera movement, light contrast, and sound should progress from open to climax to release, and which single clip breaks the pattern to mark the turn.
3. Extract ordered beats and assign each beat a status: planned, current, completed, omitted, or replaced.
4. Group beats into scenes: assign each scene one location and time envelope, one arc_position, canonical anchor_source references, max_chain_depth (default 2), and an audio plan.
5. Divide each scene into generation-sized clips using the active surface budget or conservative assumption; chain clips from accepted footage only inside a scene, and open every scene from canonical references.
6. Give every clip one narrative job, one felt_intent - a single line naming what the viewer should feel or notice, the directing engine's intention made persistent in state - and one completed endpoint.
7. Persist the complete lane record in authoring_state. A narrative clip keeps every canonical field, labels the non-transferable detail source_bound only with an exact source locator or authored_choice with a null source, adds genuinely unequal value before/after, and stores exact visible or audible prompt_carriers. Its value_before must equal the nearest narrative ancestor's value_after, walking across utility inserts. A non_narrative clip stores exactly utility_intent and non_narrative_refusal; it never fabricates psychology.
8. Define planned opening state, planned ending state, continuity locks, allowed changes, and extension-friendly handoff requirements.
9. Store later clips as provisional intent cards, not final prompts.
10. Compile only the first unresolved clip prompt from the current clip contract. Emit prompt_carriers, never the internal labels or explanations.
11. After generation, require the clip or final frame, record observed start/end state, reconcile physical and dramatic handoffs, and only then compile the next prompt.