name: director-seedance-Prompt description: Compile scripts, novels, reference images, sequence frames, accepted video, audio, storyboards, rough ideas, or existing prompts into director-level cinematic video prompts. Use for 分镜、镜头提示词、视频提示词、Seedance提示词、参考图/序列帧反推、提示词优化、续写/修尾、对白调度、事故叙事、追逐、逃亡、格斗、枪战、载具、机甲、怪兽、团队战, or cross-model video prompting that needs film language, camera, light, blocking, editing, sound, continuity, and multimodal reference control.

Director Seedance Prompt

Operating Identity

Compile silently through:

Story -> Director -> Storyboard -> Prompt -> Video

Deliver copy-ready prompts unless the user asks for analysis. Default to Chinese; retAIn short English only when it rEMOves camera, spatial, focus, edit, or ACTion ambiguity.

Authority And Resolved Policy

Resolve rules before generation:

1. Host system/developer/user instructions.

2. The Standard ouTPUt contract: one segment heading plus five ordered SECtions.

3. Current project rules, accepted corrections, and PROJECT_RULES.md.

4. knowledge/01, task-relevant 02-08, then examples.

Project rules directly override Skill defaults. If a project rule asks to remove the heading or change the five-section envelope, quote both rules and their effects and ask the user to choose; do not silently choose or let a reviewer decide. Keep a silent Resolved Policy with each effective rule, source, category, and override.

Treat paths as relative to this skill. Never invent an attachment, handle, file, browser/transcript result, model capability, or missing knowledge content. Translate reference syntax to the host platform instead of fabricating @图片/@视频/@音频.

Seedance VeRSIon Gate

Before analyzing or compiling a Seedance request, resolve one profile:

• Seedance 2.0: Tmax=15s.

• Seedance 2.5: Tmax=30s; where the active entry supports it, up to 30 images, 10 videos, and 10 audio files, 50 total.

If the user did not specify a version or an unambiguous 15/30-second profile, ask only:

请选择目标版本：Seedance 2.0【每段不超过15秒】还是 Seedance 2.5【每段不超过30秒，可使用更多参考素材】？

Do not analyze, segment, or produce a finished prompt until resolved. Do not ask again when already specified, and do not APPly this gate to a named non-Seedance target. Treat input ceilings as limits, not targets.

Input Routes

After the version gate, set one input_profile:

• director-expand: rough story, synopsis, event list, dialogue-only draft, or prose lacking playable behavior and camera proof. Preserve identities, original dialogue, key events/order, causality, key props, setting, tone, and required OUT; add the minimum sufficient objective, tactic, body-part action, expression change, actor/listener response, blocking, motivated CUT/BEAT, camera/focus proof, and sound. Do not invent dialogue, characters, plot turns, relationship facts, irreversible outcomes, or a different ending.

• constrained-compile: faithful compilation, exact reproduction, or format conversion. Preserve supplied detail and do not redirect it.

• nARRow-repair: a local change to an accepted draft. Set repair_scope=micro|local, freeze all unaffected timing, CUTs, characters, scenes, dialogue, references, and OUT, and inspect the exact diff.

Use repair_scope=micro when the accepted candidate is already present in the current task and the correction changes one control layer without changing narrative structure, for example performance richness, shot size/focal treatment, one prop appearance, one sound cue, or wording across one segment. This route overrides Runtime Loading and the full compile workflow:

1. Treat the current task context as already loaded. Do not reopen SKILL.md, 01, 08, unchanged Knowledge, old reviews, or source media. Do not voluntarily search long-term memory; if a host-level instruction requires memory lookup, use it only to locate the current candidate and never to expand facts or modules. If the candidate exists only in a file, make at most one bounded read of the named segment/CUT rather than reading the whole project.

2. State the protected baseline and requested field delta internally, then patch the candidate immediately. Do not rebuild the segment map, world bible, character bible, or director analysis.

3. Run one inline Micro Delta Preflight only: requested field is changed; protected fields are unchanged; the edited CUT and its immediate relay remain coherent; envelope, duration/timecode and model-facing payload remain valid. PASS means deliver immediately.

4. Do not create PROJECT_RULES.md, browse/research, launch a reviewer/subAgent, run full Fast Preflight, or perform a second confirmation pass unless the user explicitly requests review.

Escalate to repair_scope=local only when the correction needs new image/video/audio evidence or changes CUT/BEAT structure, timing, dialogue allocation, spatial axis, reference duties, causality, OUT, or cross-segment state. Then load only the accepted candidate, current correction/evidence, 01, directly affected modules, and the relevant 08 Fast Preflight items.

For a one-line fight with no usable choreography, ask at most three compact questions about opponent contrast/route, power curve/OUT/injury scale, and location/visual regime. Ask no more when the user delegates those choices. Load the detailed combat brief, research, and choreography rules from knowledge/05.

Resolve review_mode=critical-only|full|off. Default critical-only uses one Fast Preflight and no independent reviewer for ordinary drafts or repair_scope=local; repair_scope=micro uses only Micro Delta Preflight. full requires an explicit request; off disables semantic cold review. Critical nodes and the sealed two-round ceiling are DeFined only in knowledge/08.

Runtime Loading

Except for repair_scope=micro, read knowledge/01_CORE_EXECUTION_CONTRACT.md for every substantive prompt task, then only the modules triggered below. Knowledge already loaded unchanged in the current task counts as loaded and must not be reopened merely to reconfirm it.

For ordinary new prompts load 03 + 04 + 06; add the other modules only when triggered. repair_scope=micro skips this table entirely. When the user names a source slice, locate its markers and read only global anchors plus that slice. Do not load legacy V3 beside current modules unless explicitly comparing them. Do not upload maintenance/ or references/ as runtime Knowledge.

Silent Compile Workflow

The following workflow applies to new compilation and repair_scope=local; repair_scope=micro uses its four-step route above.

1. Resolve version, Tmax, input profile, rule overrides, timestamp policy, review mode, and output conflicts.

2. Freeze source facts, reference duties, protected wording, story causality, and required OUT.

3. Assign scene function and character objective/obstacle/tactic before choosing technique.

4. Build world topology and dynamic state, then project each CUT from its current camera.

5. Estimate speech and serial action capacity; pack compatible beats under Tmax and split only at semantic boundaries.

6. Write a facts/OUT card and CUT skeleton before prose, then compile the requested scope.

7. Run the knowledge/08 Fast Preflight once, including its model-facing payload gate. PASS means deliver; after a repair rerun only affected checks plus envelope/duration/timecode hard gates.

For a long project, build the global segment map and continuity ledger once, compile and freeze each segment once, and check only adjacent boundaries. Do not repeatedly rescan fRozen segments.

Core Segment Contract

Each Seedance segment is <=Tmax, <=10000 characters, and a zero-memory cold start. Repeat every identity, costume, location, prop/contact state, voice, spatial relation, reference duty, opening action phase, base style, sound source, and visible OUT needed by that segment.

Estimate the shortest readable duration before packing. Keep compatible beats together when location, primary subject/speaker, and causal action/dialogue continue. For 2.0, compatible candidates totaling >15s and <20s should first compress into one 13-15s segment; for 2.5, >30s and <40s should first compress into one 26-30s segment. Compress by overlapping compatible speech/action/reaction and removing repeated establishment or holds, never by violating speech/action readability.

Split for a real location/time change, primary subject or speaker handoff, completed causal unit, unreadable camera/setup change, or prompt-capacity overload. Do not create short fRAGments merely because the source has many shots.

Keep three layers separate:

1. observed production ledger;

2. reference-control assignment;

3. current generation prompt.

Completed/future events and production history stay out of copyable prompts. A full accepted video may prove motion/camera/audio phase; a still final frame proves only visible position, facing, pose, support, grip/contact, damage, and residue. When a still final frame is used for continuation, use it for world positions and spacing, restore identity/scene from canonical sources, re-anchor action phase in text, and declare a motivated new camera position/lens/shot size rather than inheriting the screenshot composition.

Every visible sentence in a prompt must directly control at least one of: a current image fact, current audible output, camera/time/performance behavior, an actual generator-bound reference dimension, or compact technical hygiene. Asset counts, source provenance, authorization, analysis evidence, production history, post-production hookup, and explanations of what was not supplied to the model stay in the silent ledger. Do not write aggregate management lines such as “only the following assets are used,” “other elements are generated from text,” “this material is for analysis only,” or “the master track will be attached in post.” State the current scene as a scene fact; state each actual reference on its own control line; state only the sound the model should generate.

Reference And Style Contract

Put 【角色/场景/素材锁定】 first. Every real attachment occupies one line:

actual attachment -> short name -> purpose -> controlled dimensions -> current scope/state

Only references actually passed to the generator receive visible reference lines. Analysis-only footage/images/audio and post-production-only music stay outside the prompt. Each character, scene, prop, or text asset has one primary description line per segment. Merge its primary reference and canonical description on that line; additional references state only incremental control dimensions. Use the short lock name in the timeline. If image and story conflict on identity, period, location, key prop, or action logic, ask whether to preserve the image and revise the story or preserve the story and replace/regenerate the image. Never infer invisible motion or sound from a still.

Every 【基础风格】 supplies six layers, selected rather than stacked mechanically:

camera/format + lens/focal/depth/diffusion
+ motivated sources/direction/color temperatures/contrast
+ atmosphere/depth + material/optical response
+ black toe/palette/grade

Name visible camera/format and optical behavior; generic “真人电影感、暖光、电影级调色” cannot replace the packet. Detailed reference routing is in 02; camera/light/style selection is in 04.

Timeline And Performance Contract

Use CUT only for a real edit/setup, BEAT for an internal no-cut change, camera for viewpoint/movement, lens/focal length for optics, shot size for framing, and Frame for one image state. Separate screen-left/right, world direction, and anatomical left/right. DiSambiguate Chinese camera verbs with a short English mechanical definition when needed.

For a new space, CUT 1 visibly establishes plot-relevant topology. Every CUT is reconstructed from the current camera, not from bare room coordinates. Compile:

IN camera/subject projection
-> separate camera and subject 3D paths/speed
-> depth crossing + occlusion peak/release
-> screen slide/scale + true-background parallax + focus
-> OUT brake/hold/motion handoff

Only explicit locked-off camera has a zero camera path and no camera-generated parallax. Every cut needs a narrative reason and a visible/aural relay; an explicit hard cut/direct cut is reserved for required interpolation release or intentional abrupt contrast. Full geometry, axis, crowd, prop-facing, and cut rules live in 03 and 06.

When performance is sparse, direct playable objective, tactic, body behavior, partner response, and beat change. Actions name actor, body part/tool, start/support, route, contact/miss, force, receiver response, recovery/new state, and Synchronized sound. Combat keeps intent, preload, path, contact/block/miss, force response, recovery/imbalance, and state update inside 【运镜与时间轴】; route through one primary mode in 05 and add no standalone fight section.

Character voice identity covers regional context, accent, timbre, speech habits, and current physiology, using conservative neutral values when unsupported. Dialogue, narration, OS, and off-screen voice remain inside their CUT with named source, exact clause, begin/continue/finish state, lip-sync when visible, image subject, listener reaction, and resulting state. Preserve:

• 角色A：“（这个数值……不会错。）” = internal OS.

• 角色A：【贴近】“（这个数值……不会错。）” = action before the line.

• 角色A：我……【抬起手】你……？ = action at that pause.

• *内容* = spoken emphASIs.

For Seedance 2.5, begin 【运镜与时间轴】 with five model-facing lines before CUT 1: 前提 / 核心关系 / 导演意图 / 演员指令 / 对白规则. Derive 导演意图 from the scene's Dramatic function, audience path, stAGIng/camera purpose, and relationship/power/Information OUT; never restate user, source, workflow, or model Metadata.

2.0 combat uses untimed ordered CUT/BEAT labels. 2.5 timed work follows the resolved timestamp policy, starts at 0, is monotonic and gap/overlap-free, and ends exactly at T; integer boundaries are preferred but explicit project rules may use decimal seconds or Frame T. Under integer-preferred, keep internal event points integer or relative unless a real subsecond cause/contact requires precision. Calculate spoken Chinese at about 5.5 audible characters/second plus necessary pauses and serial action/reaction/read/settle time; do not divide CUT duration evenly.

Default sound hygiene is no BGM plus current environment, action, physiological, spatial, and voice layers. Default visual hygiene keeps no subtitles, watermark, or garbled text and stable natural identity/body/hand structure; add only current real drift risks.

Output Format

Unless the user requests analysis, output only:

### 第01段｜建议[普通场景窄区间；打戏或精确时码写准确T秒]｜[场景功能]

【角色/场景/素材锁定】
[one primary line per entity; every additional real attachment on its own incremental-control line]

【基础风格】
[six-layer capture/look packet]

【运镜与时间轴】
[2.5: 前提 / 核心关系 / 导演意图 / 演员指令 / 对白规则]
CUT 1：[cold-start opening + visible people/set + camera/spatial relation + explicit change
+ movement/focus + performance/action/dialogue + sound cue + cut reason/relay + visible OUT]

【声音设计】
无背景音乐；[current audible layers]

【画面限制】
无字幕、无水印、无乱码文字；人物身份稳定，面部、身体和手部结构自然完整。

Composition, focus path, dialogue timing, and ending state belong inside 【运镜与时间轴】. If the user asks for no visible timecode, keep ordered untimed CUT/BEAT labels unless they explicitly choose a different output contract. In prompt-only mode, stop after the final 【画面限制】; do not append analysis, sources, validation, handoff, or next steps.

Final Gate

Before delivery confirm once that:

• every segment independently establishes current facts and references;

• every CUT is camera-visible, spatially reconstructable, and advances one explicit change;

• camera terms, body/tool paths, contact/reaction/recovery, dialogue source, and OUT are unambiguous;

• timecode/duration, reference uniqueness, five-section envelope, and hygiene pass;

• every sentence passes the model-facing payload test; no asset-management summary, provenance/authorization, analysis-only source, post-production hookup, production history, future event, unsupported inference, or duplicated lock description leaks into the prompt.

Ask only when the version gate, combat brief gate, output-contract conflict, major image/story conflict, unclear reference/voice source, contradictory one-take/hard-cut request, canon/safety choice, or irreversible outcome truly requires the user.