seedance-continuation

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 for seamless continuation, intentional next shots, bridge clips, tail repair, and re-anchoring after drift. A continuation prompt must be grounded in accepted footage, not only in the old plan.

Load the Director's Read, continuation-handoff, sequence-project-state, prompt-compiler, reference-transfer-contract, and continuity-qc. Load failure-atlas when the continuation failed or drift is visible. Load directing-engine so the next clip inherits the project's directorial voice and its position on the long-form spine; the look never re-rolls between clips.

Intent

The user already made something they accepted, and now they are trusting the story to continue from exactly where it really landed - not where the plan hoped it would. The soul of this skill is fidelity to what actually happened: honor the accepted footage as the only truth, refuse to invent the bridge, and ask for the real ending rather than guess it. Continuity is a promise that the film the user already has will not be quietly contradicted.

Required Input Gate

Before writing any continuation prompt, require:

• project_id;
• current clip_id;
• exact, non-empty parent_clip_id naming an accepted or accepted_with_deviation clip with an observed end state;
• scene_id, and whether the next clip stays inside the scene or crosses a scene boundary;
• full-story objective;
• final story outcome;
• next planned narrative job;
• next clip felt_intent - what the viewer should feel or notice;
• next clip directors_read_lane plus its complete canonical authoring_state: the full narrative record and exact visible or audible carriers, or the two-line utility intent and refusal;
• accepted previous clip or accepted final frame;
• observed_end_state;
• continuity locks;
• inherited directorial voice and arc position;
• exact reference registry;
• active surface or conservative surface profile.

If the source is unavailable, say: "I have the story plan, but I do not have the actual ending of the previous generation. Upload the clip or its final frame - python scripts/extract_last_frame.py <take> pulls the final frame locally - or describe exactly what is visible at the end. I should not invent the continuation state."

Once a frame or clip is attached and this client can actually open it, run the Observation Fast Path from continuation-handoff: the agent fills the observation record from what is visible and asks only about what the attachment cannot show (for a still: open motion, camera movement phase, audio phase). Never hand the sensing work back to the user when the pixels are genuinely in hand.

If the client accepts the file but cannot render it, the pixels are not in hand. Say so once, ask the user to describe the visible end state, and record it as reported: observation_confidence: low, requires_user_confirmation: true, and the unverified categories listed in uncertainties.

Do not hide this uncertainty by writing a speculative prompt.

After the source gate is satisfied and before the next prompt is compiled, classify the current clip with the Director's Read. Continuations classified in the narrative lane persist the complete canonical internal read against the observed end state, label the non-transferable detail as source_bound only with an exact source locator or authored_choice with a null source, add explicit value endpoints, and store exact prompt carriers. Observation/performance without a dramatic story turn and utility continuations use the non_narrative lane and persist exactly utility intent and non-narrative refusal in authoring_state; they do not invent psychology. Translate a narrative read into blocking, visible suppressed behavior, the chosen replacement for the refused stock move, camera endpoint, light, and sound; never paste internal labels into final generation prose.

Continuation Types

seamless_continuation: same shot, same geography, same open motion, same or motivated camera continuation, and accepted previous footage as the source.

intentional_next_shot: an editorial cut is appropriate. Story continuity matters, but exact frame continuity is not promised. Do not call it seamless.

bridge_between_known_states: a defined start state and end state must be connected, often with first/last-frame generation when the active surface supports it.

repair_tail: the previous final seconds failed. Repair, edit, or regenerate the tail before continuing because continuing from a failed tail amplifies the error.

reanchor_after_drift: identity, detail, geography, motion, audio, or world continuity degraded. Return to canonical identity, the strongest accepted final frame, a stable source clip, or a new intentional shot using canonical references.

Scene Boundary Rule

Crossing a scene boundary defaults to intentional_next_shot opening from canonical references. Do not promise seamless_continuation across a scene boundary; if the user explicitly asks for one, record the reason and treat the result as high drift risk.

Canon Rule

Accepted observed footage overrides planned state. If the plan says the subject reached the car door but the accepted clip ends two steps away, the next prompt begins two steps away. It does not replay the terminal exit, and it does not assume the subject is inside the car.

Rejected footage never updates canon and never becomes a parent source. Pixels can confirm a carrier, not authored psychology. If an accepted deviation changes the turn, preserve the historical planned contract, reconcile the nearest narrative ancestor's value_after with the successor's value_before across utility inserts, then revise the next obstacle/tactic and carriers before prompt compilation.

Track extension_depth as consecutive output-sourced generations since the last canonical re-anchor; it resets to 0 when a clip opens from canonical references. At the scene's max_chain_depth (default 2, hard ceiling 3), re-anchor by schedule instead of extending again. Visible drift before the cap is an immediate reanchor_after_drift.

Output Contract

Return:

1. Continuation type.
2. Source evidence used.
3. Observed end state.
4. Next clip contract, including the internal authoring handoff when applicable.
5. Intent echo: one line - "this clip exists so the viewer feels X" - confirmed before generation spends money.
6. Continuity locks and allowed changes.
7. Completed beats to exclude.
8. Reserved future beats to exclude.
9. Final natural-language Seedance prompt for the current clip only; compile the exact visible or audible carriers, never the internal authoring labels or explanations.
10. Updated Project State Capsule or a request for missing source evidence.