creator-local-reference

# Local Reference Craft This is tool knowledge for Creator, not a new entry workflow, production script or mandatory stage. Read the commissioned outcome, current script/shot, relevant cards, actual configuration and constraints. Choose methods by what must become understandable; no fixed geometry DSL, scene schema, mandatory template or prescribed skill chain. Creator owns local visual design, editable scenes, drawings, renders and previews. Storyboarder still owns shots, camera/action intent and timing; Scriptwriter owns script and inventory; Director coordinates cross-owner changes and independent acceptance. A render revealing an impossible action is evidence for that owner, not authority to rewrite the shot. Preserve existing visual identity and untouched project materials. ## Output Paths And Publication Follow [project layout](../_meta/rules/project-layout.md#task-reference-versions) for new work. Shared asset sources use `references/assets/<category>/<asset-name>/`; canonical cards and `assets/images/` identity PNGs keep their paths. Task references start at `references/epNN/tasks/taskNN/v001/`, containing needed editable/rebuild dependencies in `source/` and the complete `clean.mp4`, `caption.mp4`, `PLAN.json` delivery. Later versions are sibling `v002`, etc.; use no root `current`/`archive` or nested `finalfix` tree. Existing explicit paths remain supported without migration. Keep task handoff and optional candidate manifest at `story/work/epNN/tasks/taskNN/v001/{handoff.md,candidate-input.json}`, matching the reference version. Jobs, temporary invocation inputs, diagnostics and results belong in scoped work; code, fonts and retained partial video needed for rebuilding belong in references. Create only needed files. Reuse shared dependencies explicitly, preserve bound versions' dependencies and avoid copying all shared assets per task. Director supplies exact write paths and active reader/writer boundaries; material selection remains Creator's. Revise unpublished, unbound versions in place only without active readers. Changes to adopted, review-bound or submission-bound versions use a new sibling version. Coordinate version-name conflicts and new paths with Director through existing handoffs, without locks or an index. The canonical manifest selects current references explicitly; version maximum and mtime never select them. ## Choose The Medium - Still or 2D drawing: silhouette, color grouping, typography, a layout or one decisive pose. - 2.5D: layered artwork with limited depth/parallax when full geometry adds no useful evidence. - 3D stills: topology, scale, occlusion, lighting or multiple views of one space. - Local video: default to lightweight animatics/mixed references for whole-body movement/positions, camera/framing/scale/moves, occlusion/reveals, lighting, cuts and clock. Creator writes fine actions completely in final prompt. Completeness belongs to the final task MP4 timeline, not to SVG coverage. Creator chooses materials per shot: suitable existing asset PNGs, layered images, limited 2D animation, necessary 3D and existing video clips may be combined. SVG is one optional preferred material for editable planar compositions/poses, not a requirement for every shot or the whole group. Reuse suitable materials rather than rebuilding them for tool uniformity; SVG tooling supplies components, not the only full-task production path. Use the least elaborate medium that communicates the actual need. Asset previsuals remain optional for shape/topology, UI or multiview evidence that prose/existing references cannot adequately supply; text2image suits identity roots without required references. Preserve user-fixed style/settings and the selected provider's required outputs. Diagnose missing tools rather than building elaborate workarounds. In the existing handoff, briefly explain what can be reused, which necessary controls remain missing, and the simplest suitable material choice based on the fixed environment report. Choose per shot or coherent portion; this is a short craft rationale, not a new table, plan artifact or approval step. Treat the commission as an outcome with source and delivery boundaries. Preserve user-fixed tools and explicitly chosen local implementations where applicable; select the remaining materials yourself rather than assuming a whole-group `.blend` or CUDA route. Separate static spatial material from temporal composition. A fixed composition may use one SVG/PNG or one Blender-rendered frame, held for the source interval with timed UI/visibility layers where needed. Animate declared camera motion, overall trajectories and reveals with an appropriate minimal method; fine contact and mechanism actions belong in complete final prompt. Holds cannot replace those declared changing controls or alter shots, dialogue, cuts or the task clock. Assemble all portions into complete clean and caption review MP4s. For local video, apply the full [reference authority and omission rules](../_meta/rules/visual-prompt-craft-common.md#粗模控制与外观依据分离). 2D references control composition, positions, beats, cuts and overall trajectories. Holds, rigid translation, fixed proxy poses and limited animation are valid abstractions. State adopted controls in actual `ref.use` and final `manifest.prompt`; preserve complete contact ownership and fine action timing in final prose, with reveal order and applicable spatial/timing controls in media. Omitted hands/contact details are not floating failures or unknowns. Repair concrete conflicts within declared controls; no hypothetical-imitation rework or model-effect guarantee. When suitable reusable materials and lightweight 2D cannot express declared complex whole-object rotation, camera tracking or depth occlusion, add minimal native Blender (`bpy`) or other needed 3D to the composite. Source fine contact does not require 3D action construction. Preserve source camera/action intent. Limit new work to unresolved controls, then assemble one complete task MP4 covering every consecutive member shot; partial clips or PNGs alone never fulfill delivery. Creator always follows official prompt guidance by writing complete source-required action sequences in final prompt, with or without reference limbs/wings or locally shown fine action. Include necessary subject/part ownership, preparation/execution/recovery, contact changes and ordering; invent no actions or per-frame quotas. Official advice 7 prefers limbless rough models; its limb/wing emphasis adds imitation-risk guidance, not a completeness condition. SVD rigid BODYBOX covers whole-object staging/movement, camera, transitions, space, light and clock without arms, palms, wings or fine mechanism animation; mechanism simplification is SVD's boundary. Remove signals from existing rough animation before reuse, never shorten source action prose. Retained limbs/wings need risk handling even for spatial-only use. Independent action references keep their selected interval checks without compulsory fine rigs; every purpose preserves final source semantics. Suitable detailed models may adopt existing structure/materials or only space without copying actions. Declare dimensions and intervals; detail grants no whole-film authority. Environment/prop geometry preserves adopted spatial, adjacency and passage-topology controls, not automatic final appearance. Dedicated static shapes, including fine limbs/mechanical hands, retain their declared shape/structure purpose. Every final prompt still preserves complete source action regardless of purpose or shown detail. Rough-video simplification removes local signals without reducing prose or regenerating identity PNGs. Repair concrete conflicts, not every hypothetical behavior; spatial-only use does not waive retained rough-limb imitation risks. ## Saved Episode Media Specification Formal selected clean MP4s default to the chosen final video ratio/resolution; an explicit user local override takes precedence for local references. Verify actual provider pixel dimensions and their basis: 16:9 + 720p commonly means 1280×720, not a universal ratio mapping or short-edge rule. Resolve unknown tiers from evidence rather than guessing. Use one CFR fps throughout the episode. Honor the user's explicit fps; otherwise Creator chooses once for compatibility with canonical source clocks. An unset fps is not 30 or an example value. Before the earliest formal reference production, send actual dimensions/fps and their final-settings or explicit-local-override basis to Director for the bounded config owner to save. Read the canonical SVD_CONFIG section `## 本地参考 epNN` and its exact episode-qualified keys `epNN 本地参考宽度`, `epNN 本地参考高度`, `epNN 本地参考fps`; source/render/export parameters consume these saved values. Low-resolution drafts are valid authoring aids. Final selected clean media must match the saved spec. Inspect reusable media first; retain suitable outputs and repair only incompatible exports/conversions in scope, preserving camera/framing, ratio and canonical clock. Metadata changes alone cannot establish compliance. Missing saved spec goes to the bounded config owner, separately from environment-report recovery; add no initialization probes or automatic historical migration. Config changes use scoped review compatibility assessment, not blind fingerprint refresh. Before delivery, run the [explicit selected-media check](tools.md#saved-spec-and-selected-media-check) on a manifest-declared complete clean MP4. Check caption picture width/height, fps and duration against that clean; total caption height adds the band. Technical diagnostics support existing independent review, without a new kind, schema, automatic gate or extra round. Saved dimensions are positive even pixels, with width >= 64 for the media tools. Fps may be a positive integer, decimal or fraction; canonical task duration × fps must yield an integer number of frames. Reviews fingerprint the whole config, so coordinate authorized early saving and stable handoff rather than writing it during active evidence reads. The CFR probe's 120-second execution timeout is not a clip-duration limit. ## Direct Tool Work For coarse proxies, give actors and important props contrasting placeholder colors and values against the environment/background; keep actor/color/shape mappings stable across relevant shots. Simple surfaces and lighting preserve needed silhouettes, occlusion and overall spatial relationships without blown whites or crushed blacks. Preserve intentional concealment and reveal timing; distinguishing actors does not authorize exposing secrets. Choose scene-appropriate colors, not a universal palette or moral code. Detailed media retains adopted structure/materials under its actual purpose rather than requiring proxy recoloring. For the Director-coordinated initialization commission, use [Environment Check](tools.md#environment-check) to cover every supported local route once and save the current Markdown report at project-relative `story/work/shared/environment/environment.md`. In later work, read that fixed file yourself as a normal stable read dependency and choose from verified capabilities; Director coordinates writer ownership/stability rather than repeatedly forwarding its contents or path. If missing, request one coordinated initialization/recovery completion, not a full check by each child. Actual faults, known environment changes or newly needed unverified capabilities justify targeted updates through Director; reuse unaffected results without routine retests, TTL or version scans. Keep the report out of manifest.sources and default review semantic inputs, with no task-local copies. Unavailable routes affect only dependent work; config-only calls stay read-only. Before every image read or operation, follow [visual context and preview rules](../_meta/rules/visual-context.md): a fresh task, minimal necessary images, thumbnail-first, and text/file handoff rather than resuming an image-heavy task. This includes rendering, author inspection and each revision. Delegate directly when supported; after unavailable Task or confirmed depth rejection, reuse that finding and request top-level Director/main relay, returning actual results to the original specialist or review coordinator. Never Read an original image directly. Consult [SVG, Blender and FFmpeg knowledge](tools.md) as needed. Write task-specific drawing or `bpy` scripts in the STORY PROJECT's `references/`, not the plugin repository. Keep actual editable SVG/layered originals, scripts, fonts and other imported inputs, plus `.blend` when used; declare these dependencies in `sources`, never uploads. Typed reference media remains PNG/MP4. Inspect existing scripts before executing them. Use small manual Write/Edit/apply_patch operations (at most 2000 characters each); rendered binaries come from the tools. Optimize the author/inspect/revise cycle: reuse backgrounds, images, clips and independently editable objects; check necessary compositions and overall relationships early. For concrete failures, reconsider the method: static layers, limited animation, local 3D or mixed components may better preserve controls. Update affected sources, selected media, use and prompt together; source redesign stays with its owner. Creator chooses short drafts for unresolved declared motion/interpolation; contact drafts apply only to explicitly adopted action references. No new gate, compulsory review round or per-iteration permission follows. Keep visual operations fresh; early checks do not replace selected-media inspection or the complete task MP4. Local animation exports MP4 without persistent full-frame sequences or per-frame process setup. Per-shot clips plus FFmpeg concat are valid; preserve current durations, dialogue and cuts in one complete group output. For Blender use native batch rendering on a proven appropriate GPU path; explain unavailable/unsuitable GPU paths and tested CPU fallback before heavy work. See [direct MP4 animation](tools.md#direct-mp4-animation). The goal is fewer coupled revisions, not merely fewer render seconds or lower fidelity. Inspect the resulting MP4 in a fresh visual task, extracting only needed frames through FFmpeg and the shared preview helper. Retain temporal evidence and sampling limits for declared controls; unknown requires a necessary gap within adopted controls, not omitted fine action. Static asset PNGs remain valid with the same helper. Hand off findings for edits/renders or further inspection in new tasks. Process success is not visual success. Check local video framing, scale, layout, whole-object trajectories, camera, occlusion/reveals, lighting and clock under declared controls; escalate unresolved source-design conflicts to the owner. Missing hands are not missing support evidence. Static shapes retain commissioned geometry checks. Retain editable sources, with no universal recipe or iteration count. Image sequences require an explicit user need, not a default performance, inspection or recovery workaround. Inspect actual selected references under fresh visual-task rules. Apply [whole-frame mapping](../_meta/rules/visual-prompt-craft-common.md#全画面粗代理映射) in the final prompt's opening and each reference's actual `use`. Cover every distinguishable coarse object/class, including environment components, floors, walls, background, furniture and items without asset images. Group repeated objects of one class without omitting visual categories; no per-mesh quota or unrelated unseen objects. Remove internal trajectories, coordinates, camera cones and debug labels from uploads; retain source-intended film text under [shared craft](../_meta/rules/transition-craft.md). Fix concrete unreadable controls in affected targets. Static assets and detailed media retain declared appearance/structure/material purposes. Asset-card integration uses the optional [local reference contract](../_meta/rules/local-reference.md). On animation reuse, inspect the evaluated baseline and clocks for adopted controls, preserving intentional animation. [Contact checks](tools.md#animation-reuse-and-contact) apply only to explicitly adopted action references, not source contact alone. For missing declared-control evidence, verify render inclusion, visibility and dimensions [before camera search](tools.md#visibility-before-camera-search). Escalate infeasible source relationships to their owner. For each coarse object/class, write actual color/shape/location identifier -> final object -> adopted spatial envelope (position, screen occupancy, scale, distance, orientation, route) -> source-established final shape/structure, softness/rigidity, material surface/light response and source-action changes. Explicitly bind real uploaded identity images; describe imageless objects from established source art direction. Missing necessary design returns to its owner, not invented assets/tokens or mandatory new cards/images. Preserve camera/light/reveal/cut controls; coarse envelopes do not prescribe final box silhouettes, proportions, topology, rigidity or materials. Dedicated shape/topology and fine refs retain their declared structural/material purposes. Global stable mappings provide a reusable basis, not a once-only rule. Locally restate or fully introduce the rough proxy's final identity/shape/material when it is the main object or attention focus, when coarse appearance may be misinherited, or when expression needs it. Repetition is not limited to new objects/changes; no ban on per-segment repetition or count limit. Clarify new objects, scene ambiguity, identity switches, shape/material states or changed purpose. Avoid mechanically copying irrelevant whole-object lists; preserve purposeful emphasis, complete actions and state changes. Each explicit video use still states concrete source interval, output task interval and adopted dimensions; object introductions are need-based. Author every use for actual content and revise it with the final prompt. Official advice supplies material correspondence, story progression and useful restatement; whole-object coverage and locally explicit intervals are SVD refinements, not official mandates. ## Timed Audiovisual Rehearsal For books, screens, photos and instruments, preserve the reader/operator or recipient, usable face and plausible actor view alongside camera readability; see [interaction/viewpoint relations](../storyboarder-storyboard/camera-language.md#视线与动作接点). Do not default to surface `look_at(camera)`; use compatible framing, plausible tilt or owner-coordinated cuts. Coarse actor/eye positions and operating-region/prop layout suffice for spatial control; holding/support details belong in final prompt. Judge transparent displays, deliberate showing and levitation by intent, not a universal same-side/dot-product gate. Use cheap rehearsal for concrete spatial/timing questions: can a reveal be read before the cut, or can the camera find the prop without losing the reaction? Start with rough BOX camera/staging and [camera-language knowledge](../storyboarder-storyboard/camera-language.md). Refine unresolved declared controls such as prop scale, occlusion or focus depth; source fine contact does not require local proxies or performance animation. Preserve [beat meaning](../_meta/rules/audiovisual-craft.md): initial state, necessary visible/audible evidence, ordering versus credible overlap and attention. Use expert estimates of dialogue, listening, breath and reaction with optional [fake audio timing](tools.md#fake-audio-timing). Distinct tones can expose pauses, simultaneous speech and sound bridges against rough cuts; they are not TTS, semantic inference, intelligibility or performance acceptance. Silence can be occupied by listening or reaction, not spare time to fill. Neither a tone file nor a successful render proves the scene works. Report the observed conflict and its effect before offering optional fixes. When timing or coverage needs redesign, Director coordinates owners under the [original episode budget](../director-orchestrate/SKILL.md#集总时长责任): use its confirmed +/-10% creatively without per-adjustment permission, update canonical script/storyboard first, then affected references/manifest. Exact user targets and scope still bind; do not ratchet the baseline or silently stretch reference time. Self-inspection and each visual revision still use fresh tasks and helper previews; no additional rehearsal ledger, required artifact or gate. For UI/control actions, preserve trigger, feedback and consequence in complete final prose, including completion/release/freeze. Under [event-state guidance](tools.md#ui-and-event-states), rough media supplies operating regions, overall relationships and source-timed UI results, not fine contact/mechanism evidence. Inspect actual displayed states; pixel/duration checks prove only measured properties. Segmented repairs require full-composite and relevant-join inspection for unintended noise, light/color, motion and audio jumps. Repair contradictory declared controls before prompt handoff. For lightweight animatics/mixed references, deliver both the complete clean task MP4 and a same-source complete [caption review MP4](tools.md#caption-review-mp4), using `previs-preview.py` with `--timecode` and the existing PLAN format. This delivery applies regardless of material mix. Include every source dialogue/narration utterance verbatim in its corresponding task-time windows, including source-intended continuation across cuts. Completeness means full utterance/window coverage, not text throughout silence. Preserve clean duration, frame rate and cuts; no retiming, TTS or fake-audio generation is required. Actually export and return both videos and PLAN. In the complete review MP4, mark every actual internal cut and its task-global time through existing display-only PLAN segments; the tool does not infer cuts. A single-shot task without internal cuts needs no invented markers; report that fact. Keep internal markers out of clean media and final prompt, and keep the caption version outside uploads; source-intended film text stays in clean. The pure-text owner never views caption media; any viewing uses fresh scoped tasks/helper thumbnails. Captions support human timing review, not independent acceptance or clean-picture proof. Add no schema, gate or extra review round; report tool limitations through Director/main rather than delivering only shortened clips. ## Episode Caption Delivery After all current-episode materials and existing independent reviews are ready and dependencies stable, Director automatically commissions one Creator to assemble the episode caption preview within the production scope. Keep every complete task clean + caption MP4 and PLAN; also deliver `references/epNN/episode-previs/review.mp4`. Series delivery happens once per completed episode, not all-series. A local/partial commission does not expand to an episode assembly. Follow the [episode tool contract](tools.md#episode-caption-review-mp4); use the confirmed interface before execution. Write `story/work/epNN/episode-previs/parts.json` as an ordered array of `{video,plan}`, one entry per group in canonical source-task order. Explicitly choose one manifest-declared complete clean MP4 per group and its actual caption PLAN. Do not infer PLAN from `sources`, JSON extensions or filename conventions. This is disposable invocation input, not a ledger, editable timing authority or new manifest field. Read/write only Director's bounded stable dependencies and assigned paths; escalate missing mappings or conflicts before dependent work. Assembly PLANs contain `segments` and optional `context`, with no other top-level keys. Keep verbatim dialogue/narration and every corresponding task-local window in `segments`; put scene/camera/action/performance explanations in `context`, never disguised as speech. If original PLANs mix internal labels or cues with speech, semantically prepare assembly PLANs under `references/epNN/episode-previs/`, preserving originals and every source word/window. Do not regex-delete words resembling time/cut labels. The tool derives task/SHOT labels from canonical inputs and rebases dialogue and context spans once by cumulative canonical task durations. Keep input spans task-local. Concatenate clean media first, then caption the complete episode; never concatenate burned-in caption task videos. Derive offsets from canonical durations, not rounded container lengths. Every PARTS clean video must match this episode's saved width/height/fps, with verified CFR, square pixels and real canonical duration agreement. Repair incompatible parts in scope before assembly; do not hide dimension errors with automatic maximum-canvas padding. Preserve camera/framing, ratio and clocks during repair. The caption picture keeps saved dimensions; its total height adds the band. Preserve present audio, fill absent-audio parts with equal-duration silence when needed, and keep an all-silent episode without an audio track. This adds no TTS or seamless J/L-cut guarantee. The CLI requires a new output and has no overwrite flag. For updates, Creator generates and verifies a new candidate, then safely replaces the formal file within the existing authorized scope and stable dependencies. Resolve PARTS video/plan paths from the story project root, not the mapping's directory; plugin-root paths locate the tool. Return actual stdout output, total duration, task/shot order, mapping, dimensions, fps/audio and inspection limits. Failures/partial outputs remain incomplete delivery. This episode reference is internal, not generated-video editing or an upload; it creates no formal review kind/gate and does not modify original reviews, manifests or grants. Author checks do not issue independent pass; any viewing follows fresh visual-task rules. For new formal episode delivery, Creator supplies source-based scene and applicable camera/action descriptions for each current beat in `PLAN.context`. Entries are `{shot,scene,spans,camera?,action?,performance?}`: `shot` is an integer member of that task, `scene` a nonblank string, and nonempty `spans` contain task-local `[start,end)` windows wholly within that source shot. Optional text fields must be nonblank when present. Supply performance only where the source supports it. Each window is complete with all applicable fields; nothing carries over from earlier entries. Keep the same scene name across tasks in the same scene. Combine simultaneous fields in one entry: distinct entries cannot overlap; overlapping spans of one entry display once. The episode band separates `【说明】` from `【对白/旁白】`, with the current canonical task/SHOT range and global `HH:MM:SS` clock. Uncovered context windows display `未提供`; absent camera/action together also display `运镜/动作：未提供` and produce stdout `warnings`. Legacy PLANs without context remain readable, but tool success or empty warnings do not establish complete new delivery. The tool does not infer prose or add manifest metadata. The renderer rejects explanation content exceeding six wrapped lines at any instant; this is a tool layout limit, not a creative quota. First express the current beat concisely and accurately, or split changing beats into legal source-shot windows with complete applicable fields in each. Preserve key facts and original timing; do not truncate facts, add seconds or merely split identical overlong text spans. Report unresolved delivery limits to Director. Single-task `previs-preview.py` retains its segment-based caption behavior; it does not process top-level context or derive canonical labels. ## Grouped Task References Before final prompt authorship, apply the [existing-asset selection loop](../_meta/rules/visual-prompt-craft-common.md#已有资产选用闭环). From authorized stable sources and actual reference observations, mentally or in the existing handoff align objects/classes, established identity/form/material, corresponding existing cards/images, each using shot's header and current material slots. Check relevant known assets first; reuse suitable established images, correcting unselected assets whose omission affects identity/form/material. No applicable image allows source-based description, not compulsory new cards/images, whole-library inspection/uploads or a new form registry. Group background classes as appropriate. Unsupported extra objects require media repair, not invented assets. Report missing declarations with exact objects, shots, card/image paths, source basis and impact to Director for a bounded stable Storyboarder header edit. Story/asset-list/design changes go to their owners; Creator does not edit upstream. After stable handback, rerun provider materials and use actual asset-union/local-media slots to semantically revise the whole prompt and every affected use, including existing timed bindings. Keep header -> first-use asset union -> local media; no parallel upload list, guessed appended token or submission-time replacement. Self-check final --json and obtain applicable scoped review with current fingerprints; source/binding changes do not waive gates. Every task's authorship, local video intervals, complete source actions and clean/caption/episode delivery still apply. Apply [transition and film-text craft](../_meta/rules/transition-craft.md): clean MP4 means no internal debug contamination, not no text. Source-intended UI, audience SUPERs, time/place/chapter/full-screen cards and transition imagery are valid. Default exclusion of text labels concerns internal positional/debug labels. Optional [local text guides](tools.md#film-text-and-transition-guides) preserve source words and Storyboarder's reading/cut windows in existing media/use/prompt. Standalone cards count in canonical shot/time budgets; overlays share host time and assembly adds no seconds. Script owns words/time facts; new day jumps return to that owner. Input acceptance promises no final model text accuracy. At independent TASK boundaries, strongly prefer motivated, visibly distinct camera/view/scale by default to reduce near-identical generation mismatch visibility. Apply junction intent under Authority And Delivery below. Group on motivated source cuts; similar shots may stay together. Preserve needed match/repeated compositions and keep essential uninterrupted contact/speech together when feasible within model maximum. Explain choices in the existing handoff without new permission. This is not an every-shot change, angle quota, continuity guarantee or excuse for source-conflicting identity/state/action. TASK may contain multiple shots/scenes within constraints, without compulsory acts or task=scene. Mid-motion cuts need no settling/restart. Preserve consecutive membership, durations and grants; source redesign goes through Director/owners within the original budget, never silently changing cuts or protected groups. Read [shot-inputs.md](../_meta/rules/shot-inputs.md). After shot design, group consecutive photographic shots without changing their current canonical durations, dialogue or cuts. Verify actual model maximum M and aim for task duration `ceil(0.7*M)..M`, a semantic packing target, not a mechanical lower quota. Validate actual provider task limits; unsuitable grouping returns to owners for redesign within the original episode budget, not silent stretching during assembly. Write a draft `story/episodes/{ep}/task-inputs/taskNN.json` with exactly `{shots,references}`, then finalize it as exactly `{shots,references,prompt}`. Creator authors the nonblank string `prompt` after reference video and group mapping are settled. Filename-derived task_id is stable independently of first member. Members are ordered positive safe integers consecutive in source order. Derive timing from storyboard; keep no editable duration/offset fields or duplicate grouping index. Each task needs a full-group MP4, optional PNGs, actual editable sources and explicit use/placeholder limits. GIF is unsupported. Use one coherent task reference clock for all member intervals, internal cuts, positions, trajectories and sound bridges. Static intervals may use static clips. Header identity assets form a first-use union before local media; sources do not upload. Obtain authoring materials through the selected provider's own tool, following its video guide (Dreamina: [video.md](../creator-provider-dreamina/video.md#dreamina-authoring-materials)). Generic converter CLIs use `--json STORYBOARD TASK_ID EP` for the finalized package, without generating text. Inspect task_id/shots/timeline and report continuity dependencies. Partial selected-shot scope reports full membership and additional members rather than silently expanding authority. Assembly does not authorize tasks.json preparation or submission. Follow [publication and selection rules](../_meta/rules/shot-inputs.md): keep published JSON valid after each bounded edit; optionally stage the existing draft/final shape in the version's work `candidate-input.json`, validating the complete document before authorized atomic replacement. Once Director has established stable scope/dependencies and canonical write ownership, promote the chosen package to `story/episodes/epNN/task-inputs/taskNN.json` with matching prompt, uses, sources, media and actual source timings. Self-check that published final input, then hand the canonical manifest to a fresh independent Reviewer. Work candidates are not review targets and their checks do not establish final readiness. Preserve existing shapes, gates and protected submitted records. ## Authority And Delivery For selected transitions, express A's end state -> trigger -> transition process -> B's start state with actual subjects/views, necessary direction, shot scale, timing and sound. Map source windows to task time in media/use/prompt. A hard cut switches instantly at its cut point; it needs no invented intermediate movement. Distinguish sequential fade-to-black/appearance from simultaneous cross-dissolve. Preserve established choices unless a concrete conflict goes to owners; do not add transition seconds or impose smoothness, stops or dissolves. See [transition process](../_meta/rules/transition-craft.md#转场过程表达). Apply [source-intended editorial relationships](../_meta/rules/shot-inputs.md#reference-authority) to each junction. Within the same continuous event, carry necessary compatible action progress, possession/contact, space and sound continuity into selected controls and final prompt; do not invent a stop, end pose, pause or restart, or require identical frames. At scene/act/time/place jumps, preserve meaningful causal, emotional, information, thematic contrast or parallel relationships and viewer orientation appropriate to the reveal. Do not force same-position, continuing action/sound or a bridge scene. Source-supported mystery, abruptness and hard cuts need no smoothing or immediate explanation. Same-episode underlying identity/world facts remain consistent except source-justified intentional changes. Before handoff, inspect relevant junctions as coherent tail/head windows of the actual selected clean MP4s in fresh visual tasks, alongside final prompt intent. Report actual audio present/heard separately from planned J/L bridges, internal fake timing and independently generated audio limitations. Source dialogue remains once at its origin with explicit continuation; do not restart it in each task. Pass necessary neighbor paths and coverage limits to the independent text owner; missing neighbor evidence is scoped unknown, not authority to generate that neighbor or auto-edit final videos. Before authoring performance, read [source trigger-to-response and task-time integration](../_meta/rules/audiovisual-craft.md#从触发到人物反应). Preserve complete source-defined responses in final prompt and their applicable spatial/timing controls in BOX staging. Missing story facts go to Scriptwriter; camera/coverage redesign goes to Storyboarder through Director. Preserve the script/shot's necessary screen or page facts, exact wording when meaning depends on it, key action/results/reactions and the order in which they become understandable. A visible clue or a readable surface alone does not establish its interpretation or justify a choice. Keep those facts in the final authored prompt and their applicable controls in references; internal labels and BOX placeholders are not substitutes. Report missing story support to Scriptwriter through Director and missing framing/view/cut/read-window design to Storyboarder, without inventing evidence or giving every routine operation equal emphasis. Direct dialogue, narration and sound remain valid information carriers. Read canonical sources, provider materials/reference syntax and actual refs; pack, tokens, binding and rebasing belong to that provider. Before review, author complete task-time `manifest.prompt`: preserve narrative, exact dialogue/film text, cuts, duration and audio. Write all fine grasp/release, palm/wrist support and mechanism actions with who, which anatomical hand/part, ownership, direction, initial/intermediate/end changes, ordering/overlap and result. Bind real refs and map BOX/colors or existing limbs to identities and adopted controls; final anatomy/action follows source and identity, not unadopted proxy poses. Remove only non-film headings, internal IDs/paths and review metadata. Missing facts return to owners. Self-check actual final `--json`, then obtain fresh independent shot-input review for fidelity, completeness and integration. Drafts are not ready; submission never rewrites accepted text. Apply the shared [detailed shot prose and proxy authority rules](../_meta/rules/visual-prompt-craft-common.md#粗模控制与外观依据分离). Using actual reference inspection evidence, identify likely unintended transfer and apply [reference-specific exclusions](../_meta/rules/visual-prompt-craft-common.md#按实际参考限定非目标特征) in the relevant `use` and task-time passage of the exact `manifest.prompt` you author before review. Coordinate missing shot detail with Storyboarder. Read and apply [global mappings and inline use](../_meta/rules/visual-prompt-craft-common.md#全局映射与实际使用处的引用) while writing the final task-time passages; bind the real inputs to the source performance integrated above.