Story To Reusable Prompt

Story Approval to Video Prompts

Outcome

Keep these operations distinct:

1. determine whether the user already supplied a usable story premise; route only product/brief inputs through optional direction selection and keep that strategy work outside the story engine;
2. refine the supplied story or selected direction with character/scene settings, speech intent or draft lines, and natural sound;
3. after story approval, test duration fit and, when materially different viable camera/editing strategies exist, let the user select a compact shot-structure direction before exact timed planning;
4. jointly design timed shots and final speech in a compact audiovisual plan for review;
5. before that review, normalize production-critical character, scene, product, prop, and boundary states into a compact production bible;
6. approve the audiovisual plan and production bible together, then hand off the locked package for video-prompt compilation without expanding dialogue or rewriting the plot.

Scope

Handle:

• product, selling-point, audience, or rough-idea inputs that need a story direction before story completion;
• selected direction cards and compact upstream direction handoffs;
• rough outlines, fragmented scenes, incomplete plots, and complete final stories;
• causal repair, requested character/scene/performance design, story-stage dialogue and natural sound, and continuity;
• separation of user facts, reversible assumptions, AI-created material, and unknowns;
• post-approval beat segmentation, timed shot planning, continuity, sound, dialogue, voiceover, and AI video prompts.

Do not absorb or duplicate the full product-strategy and multi-direction methodology owned by short-video-director-thinking. Do not abstract the story into a reusable story generator or audit unrelated prompts. Do not automatically add web research, visual-reference search, reference-image boards, media generation, publishing, or fixed marketing tactics. Model-specific syntax is a replaceable downstream adapter, not a core story rule.

Single Entry and Internal Routing

Read references/workflow-routing.md when the request starts from a product without a story, spans stages, asks what happens next, or could be confused with direction, production-bible, storyboard, final-video-prompt, or reusable-Prompt work.

Accept a product brief, selected direction, outline, or story through one public flow. Internally route it as: determine entry artifact → select a direction only when necessary → resolve important story choices → integrate story, visuals, and speech intent → story review → compact audiovisual plan → plan review → locked handoff → final prompts. Stop at the requested deliverable. Do not ask users to choose an entrance or name another skill. Preserve settled content; narrow edits do not reopen unrelated settings. Speech intent belongs in the story, final wording and timing in the audiovisual plan. Do not add a separate voice-only workflow or approval round.

Treat an input as a usable story premise when it contains, explicitly or by clear implication:

• an acting subject or focal character/object;
• at least one event, problem, goal, conflict, demonstration, or meaningful action change;
• an intended outcome, reveal, result, or ending direction.

An incomplete premise still counts as a story. Missing appearance, scene, dialogue, causality, or ending detail belongs to story completion and must not trigger fresh direction ideation. A product name, feature list, target audience, desired mood, content type, or vague request such as “make a funny product video” without an event chain is a product/brief input, not a story. The presence of a product inside an outline is never by itself a reason to reopen direction strategy. Reopen direction ideation only when the user explicitly asks to compare, replace, or rethink creative directions.

Infer the narrowest internal deliverable mode, never a user-facing entrance menu:

• product_brief_to_story: obtain and approve a story direction first, then complete the selected direction into one integrated story;
• selected_direction_to_video_prompts: bind a selected-direction handoff, complete and approve the story, then continue through the requested downstream stages;
• outline_to_video_prompts: complete and approve the story, review its audiovisual plan, then compile final prompts;
• story_completion: return only the integrated story for review; no downstream work is requested;
• approved_story_to_video_prompts: skip story revision for approved/final copy; plan shots and speech, or compile an already approved audiovisual plan;
• one_shot_override: complete the story and continue to video prompts in one response only when the user explicitly asks to skip confirmation.

State and Approval Gate

Track:

ENTRY_ARTIFACT = PRODUCT_BRIEF_ONLY | SELECTED_DIRECTION | STORY_DRAFT | APPROVED_STORY | APPROVED_AV_PACKAGE
DIRECTION_STATUS = NOT_REQUIRED | WAITING_FOR_SELECTION | SELECTED
STORY_STATUS = DRAFT | WAITING_FOR_CONFIRMATION | APPROVED
AV_PLAN_STATUS = DRAFT | WAITING_FOR_CONFIRMATION | APPROVED
PRODUCTION_BIBLE_STATUS = NOT_BUILT | DRAFT | APPROVED_WITH_AV_PLAN
PIPELINE_STAGE = DIRECTION_IDEATION | DIRECTION_SELECTION | STORY_REFINEMENT | STORY_APPROVAL | DURATION_AND_STRUCTURE | AV_PLAN_AND_BIBLE_APPROVAL | OPTIONAL_REFERENCE_PREP | VIDEO_PROMPT_COMPILATION | REUSABLE_PROMPT_COMPILATION | COMPLETE
DOWNSTREAM_TARGET = NONE | AV_PLAN | VIDEO_PROMPT_SET
SPEECH_DECISION = UNASSESSED | DIALOGUE | VOICEOVER | DIALOGUE_AND_VOICEOVER | INTENTIONAL_SILENCE
DURATION_FIT_STATUS = UNASSESSED | FIT | FIT_AFTER_NON_MATERIAL_COMPRESSION | FIT_PENDING_USER_APPROVAL | DURATION_CONFLICT
GENERATION_RISK = UNASSESSED | LOW | MEDIUM | HIGH | MODEL_LIMIT_CONFLICT
SHOT_STRUCTURE_STATUS = UNASSESSED | NOT_REQUIRED | OPTIONS_READY | SELECTED
SHOT_STRUCTURE_SELECTION_MODE = USER_CHOICE | USER_LOCKED | AUTO_RECOMMEND | SINGLE_FEASIBLE

For a rough or materially incomplete story with a downstream target:

1. repair and present the complete story;
2. set STORY_STATUS: WAITING_FOR_CONFIRMATION;
3. stop before timed segmentation, final dialogue budgeting, shot prompts, or model adapters. Include supplied dialogue, necessary speech intent or brief draft lines, and sound cues, not a word-count-locked recording script.

Treat phrases such as “确认剧情”, “剧情通过”, “按这版做分镜”, or an equally clear instruction as approval. Feedback or requested changes without explicit post-revision approval are not approval: revise the smallest affected story unit and remain WAITING_FOR_CONFIRMATION. If one message gives deterministic edits and explicitly says to continue after applying them, the revised version may become APPROVED; if the edits require interpretation or new invention, present the revision and wait.

Once APPROVED, freeze the premise, event order, causal chain, reveal, resolution, user-locked details, reviewed character/scene settings, and reviewed speech/sound intent. Preserve speakers, line meaning, information-release position, language, delivery intent, and intentional silence; exact wording is fixed when explicitly locked. Shot design may clarify technical execution but may not redesign those elements or add a new event, motivation, product claim, or ending. A later material edit to a frozen element returns the story to WAITING_FOR_CONFIRMATION; non-locked wording or technical execution that preserves reviewed meaning and intent may keep it approved.

Story approval does not approve an uncreated audiovisual plan, its production bible, or AI draft wording. Set AV_PLAN_STATUS: WAITING_FOR_CONFIRMATION and PRODUCTION_BIBLE_STATUS: DRAFT when presenting them together. Clear approval applies to both artifacts and permits compilation; story-only approval does not. Do not create a separate production-bible approval round. An explicitly supplied final audiovisual plan and production bible need validation, not a redundant approval request. After joint approval, preserve exact final lines, speaking windows, pauses, silence, entity fingerprints, state deltas, and continuity boundaries except declared variable slots. A timing, line, or material production-bible change reopens only their joint review; a changed story meaning also reopens story approval. An explicit instruction to skip intermediate reviews and deliver directly may bypass pauses, never feasibility or production-bible checks.

Do not repeat story approval when the user clearly says the supplied story is final; audiovisual-plan approval remains separate. Do not infer approval from silence. For story_completion with DOWNSTREAM_TARGET: NONE, complete the story and stop without requesting approval.

Input Discipline

Read references/story-completion-contract.md before Phase 0/A for existing-setting extraction, critical missing-setting choices, edit authority, placeholders, and repair versus expression rules.

Maintain CURRENT_STORY_STATE from the latest complete draft plus subsequent explicit corrections. A complete replacement draft supersedes earlier drafts; partial feedback changes only the named elements and preserves the rest of the current story. Do not restore an older alternative character, prop, action, information source, proof action, or ending detail merely because it appeared in history.

When direction ideation is required, retain only the selected direction as the upstream creative source. Do not carry rejected alternatives, full product-analysis prose, scoring tables, or repeated recommendation rationales into story completion. Bind the selected result as a compact SELECTED_DIRECTION_HANDOFF:

SOURCE_PRODUCT_FACTS_AND_REFERENCES
SELECTED_DIRECTION_PREMISE
CORE_SELLING_POINT
TARGET_AUDIENCE_AND_MARKET
HOOK_AND_CORE_CONFLICT
CHARACTER_RELATIONSHIP_FUNCTION
SCENE_FUNCTION
CORE_ACTION_CHAIN
PRODUCT_ENTRY_AND_PROOF_ROUTE
MUST_PRESERVE_CAUSAL_BEATS
ENDING_OR_CONVERSION_FUNCTION
CLAIM_AND_EVIDENCE_BOUNDARIES
FIXED_FIELDS
FLEXIBLE_FIELDS
UNRESOLVED_DEPENDENCIES

The handoff is an abstract creative lock, not a completed story, timed plan, final dialogue, or story approval. Preserve its fixed fields and factual boundaries. Expand its flexible fields only during authorized story design. If the user directly supplies a chosen direction without this schema, extract the minimum equivalent record internally instead of forcing them to reformat it.

Classify input into:

• USER_LOCKED: explicitly required and not open to change;
• USER_SUPPLIED: directly stated story information that is not necessarily locked or approved;
• INFERRED: a reasonable, reversible interpretation;
• AI_CREATED: fictional material added for repair or within authorized design scope;
• UNKNOWN: information that cannot reliably be determined.

Before completing or optimizing the story, detect:

• USER_INPUT_CONFLICT: two current user-supplied statements assign incompatible identities, positions, actions, states, or causes;
• UNESTABLISHED_REFERENCE: a later sentence relies on an event, prop, message, clue, or information source that the current draft never establishes;
• DETAIL_EXPANSION_RISK: a generic supplied statement would become more specific through an added person, object, location, action path, message, or proof action;
• MATERIAL_COMPRESSION_CHANGE: meeting a requested duration would require adding, deleting, replacing, reordering, or reinterpreting story content rather than only tightening performance, overlap, or editing.

Use the contract's edit-authority rules: preserve settled content, execute authorized design, defer permitted unknowns as 【placeholders】, and ask only about unresolved material choices outside that authority. Contradictory facts are not placeholders. Physical plausibility does not authorize replacing a user choice.

Also capture, when relevant:

• DOWNSTREAM_TARGET
• TOTAL_DURATION
• TARGET_PLATFORM
• TARGET_MODEL
• ASPECT_RATIO
• VISUAL_PROMPT_LANGUAGE
• VOICE_LANGUAGE
• EXPLANATION_LANGUAGE
• TARGET_COUNTRY
• MARKET
• CONTENT_BOUNDARIES
• PRODUCT_PROFILE
• HERO_ACTION
• PROOF_ROUTE
• PRODUCT_STATE_PATH
• HOOK_STRATEGY
• PERSONA_STRATEGY
• SCENE_STRATEGY
• PROOF_STRATEGY
• CTA_STRATEGY
• NARRATIVE_STRUCTURE_DECISION
• METRIC_OBJECTIVE

Do not present inference or invention as user fact. Do not invent real-world prices, inventory, endorsements, certifications, product effects, or model capabilities.

When these product or strategy fields arrive from an upstream reusable-prompt specification, treat supported immutable features, the selected Hero Action, proof route, approved product state path, and user-locked strategy choices as locks. Preserve the original product information or reference identifiers as the source of truth. Clarify visible staging when needed, but do not silently replace the selected interaction, proof mechanism, state transition, Hook promise, Persona function, or scene function. Treat NARRATIVE_STRUCTURE_DECISION as abstract guidance rather than a fixed scene or shot template. Translate the strategy into production decisions without turning a metric objective into a guaranteed result.

When an upstream LOCKED_STORY_BLUEPRINT contains canonical Beat IDs, treat every canonical Beat as USER_LOCKED. Do not extract a replacement Beat list, renumber the Beats with new meanings, or let product analysis override the story. Map each final Shot and Segment to one or more canonical IDs, preserve their order, and require the union of mapped IDs to exactly equal the required set. If credible non-material packing cannot hold them, return FIT_PENDING_USER_APPROVAL when a specific material alternative exists; return DURATION_CONFLICT when the maintained locks make fitting impossible. Never replace the story with a shorter product demonstration silently.

Use minimum necessary invention for causal repair, not as a ban on requested expression or design. Work within the declared scope; preserve generic details outside it. Permission to design a character does not authorize a different ending, and permission to design a demonstration does not authorize new product capabilities.

A duration request alone authorizes meaning-preserving timing and editing decisions, not story changes. Apply separately authorized story edits within their scope; otherwise use FIT_PENDING_USER_APPROVAL to present the smallest proposed material changes. Return DURATION_CONFLICT only when maintained locks cannot fit or no accepted repair remains.

Ask at most three concise questions per round for unresolved choices that affect story meaning, important visual intent, or necessary speech language under the completion contract. Do not ask for technical voice parameters. A timed plan requires this task's total duration; when unknown, ask once or provide an explicitly requested untimed plan without word budgets. Do not inherit another story's duration or treat a framework-test default as the user's production requirement.

Core Workflow

Pre-Phase — Determine whether direction selection is needed

1. Classify the latest valid artifact using the usable-story test above; prefer the user's current artifact over older conversation material.
2. For a rough, incomplete, complete, or approved story, set DIRECTION_STATUS: NOT_REQUIRED and proceed directly to Phase 0. Do not offer alternative directions or rerun product strategy unless explicitly requested.
3. For an already selected direction, bind or extract SELECTED_DIRECTION_HANDOFF, set DIRECTION_STATUS: SELECTED, and proceed to Phase 0. Direction selection does not set STORY_STATUS: APPROVED.
4. For PRODUCT_BRIEF_ONLY, use short-video-director-thinking when available to generate and assess genuinely distinct direction options. Keep its product strategy, hook design, direction comparison, and recommendation logic in that skill rather than reproducing it here.
5. If the user asked only for directions, return that skill's direction result and stop. If the user requested the continuing production workflow, present the viable directions, set DIRECTION_STATUS: WAITING_FOR_SELECTION, and stop for a selection unless the user explicitly authorized automatic selection.
6. After selection, keep only the compact selected-direction handoff, set DIRECTION_STATUS: SELECTED, and continue to Phase 0/A. Do not repeat the rejected options or require the user to invoke the next skill manually.
7. If the direction skill is unavailable, disclose the missing optional direction stage and request a story premise or chosen direction; do not silently imitate its full methodology inside this skill.

Phase 0 — Validate the current draft

Before Phase A:

1. bind the current draft and corrections; extract existing settings using the completion contract, preserving active locks, authorized design fields, and deferred placeholders;
2. check for contradictions, unestablished references, important narrative or visual setting gaps, existing speech/silence requirements, SPEECH_DECISION, and necessary target-country speech language under the completion contract;
3. separate timing-only compression from material story changes;
4. if a material decision outside existing authority still blocks the requested work, use the contract's choice pause for missing settings or quote source phrases for conflicts; otherwise continue without a questionnaire, keeping deferred dependencies explicit.

Phase A — Complete and approve the story

1. Normalize the requested deliverable, supplied constraints, and locked details.
2. Map only the story functions actually needed: trigger, goal, obstacle, attempt, escalation, turn, proof or climax, and resolution are options, not a mandatory formula. When an upstream NARRATIVE_STRUCTURE_DECISION exists, use its required functions and repair requirements as guidance, but preserve a supplied or approved story over an unapproved structure candidate.
3. Distinguish causal repair from requested expression/design; a coherent story may still benefit from the latter.
4. Repair genuine gaps with the least new information possible; prefer timing, observation, body movement, and consequences already implied by the input.
5. After the causal story works and before presenting it for approval, set SPEECH_DECISION. Missing supplied lines means UNASSESSED, never automatic silence. Decide whether the performance calls for in-scene dialogue, voiceover, both, or intentional silence. If speech serves the immediate reaction, conflict, information release, or character goal, integrate its intent or a brief draft; if silence serves the beat better, make that an affirmative reviewed decision. Do not create a separate voice-only approval round.
6. Integrate selected character/scene details, supplied dialogue, necessary speech intent or brief draft lines, and action-grounded sound using the completion contract. Identify who addresses whom, required meaning, reveal position, and emotion; defer final wording and word budgets to Phase B. Dialogue must arise from the character's immediate perception, pressure, relationship, and action. Remove lines that narrate visible action, explain the plot to the audience, sound like generic AI prose, or turn a character into a marketer. Preserve exact-copy locks, intentional silence, and narrow scope. Resolve target-country speech language through references/localization-adapter.md before drafting target-language lines.
7. Preserve locks and ambiguity; distinguish proposed material from supplied facts. Keep product-dependent 【placeholders】 when permitted, and recheck affected dependencies after edits or slot filling.
8. Apply relevant story-quality gates, including character knowledge, scene action access, speech decision, dialogue naturalness, and localization when they are triggered.
9. Present one integrated story for user review, not separate plot, visual, and audio approvals. If DOWNSTREAM_TARGET: NONE, stop; otherwise set WAITING_FOR_CONFIRMATION and wait for explicit approval of that version.

Phase A may run only after Phase 0 has no unresolved material conflict. one_shot_override skips only the intermediate reviews the user explicitly waived; it does not bypass Phase 0, timing checks, or authority limits.

Phase B — Joint audiovisual planning and dialogue finalization

Run for AV_PLAN or VIDEO_PROMPT_SET after the story is APPROVED.

First read references/duration-fit-and-compression.md, references/shot-structure-selector.md, and its references/shot-language-library.md; then read references/dynamic-storyboard-engine.md. When the user supplies a target country/market, speech language, dialect, or regional-expression request, also read references/localization-adapter.md before speech timing. A supplied target country determines the localized speech language under that adapter; it does not by itself add speech to an approved silent story or alter the plot, cast, or location.

0. run the semantic duration preflight before committing to shots: map source units by causal, information-release, payoff, visual-proof, continuity, expressive, redundant, and user-lock functions; assign the audience tasks they must perform; try non-material screen compression first; establish the minimum feasibility envelope without assuming prose length or a fixed shot count;
1. build the minimum coverage graph, then generate two or three materially different viable shot-structure candidates only when the user requests options or the choice would materially change viewer experience, evidence, rhythm, reveal, performance continuity, or production risk. Evaluate DURATION_FIT_STATUS, GENERATION_RISK, and continuity cost separately for every candidate. Never offer an infeasible option for variety;
2. if a real choice remains, return the selector's compact candidates with one recommendation, set SHOT_STRUCTURE_STATUS: OPTIONS_READY, keep AV_PLAN_STATUS: DRAFT, and stop before exact timed shots. If the user locked a structure, asks the skill to choose, or only one route is feasible, validate or select it and continue without a redundant pause. For FIT_PENDING_USER_APPROVAL, present at most three smallest material story options with estimated savings and losses; return DURATION_CONFLICT only after credible non-material packing and shot structures are exhausted or maintained locks make fitting impossible;
3. freeze SELECTED_SHOT_STRUCTURE; identify the story type, primary hook, product-entry function, proof type, and ending function; use NOT_APPLICABLE for non-product or non-proof stories;
4. map every required source unit and audience task to the plan, then group compatible beats into economical core audiovisual blocks and actual Shots under the selected structure. Beat, Block, Shot, Panel, and Generation Segment counts remain distinct;
5. provisionally allocate action/recognition time and speaking windows within the supplied total;
6. finalize target-country lines inside those windows, preserving approved intent and exact-copy locks; make them character-specific, situation-bound, speakable aloud, and free of unrequested slogans, benefit lists, CTAs, or presenter language;
7. count words/characters and estimate delivery including pauses; jointly revise wording and local shot allocation, not one around a prematurely frozen other;
8. validate every actual cut's access or expressive purpose, the selected viewer relationship, practical capture style, actions, speech, silence, continuity, total time, source-unit/audience-task coverage, and generation risk. Surface unresolved overload rather than accelerating speech unnaturally. A duration fit with HIGH generation risk calls for safer coverage or packaging, not automatic story deletion;
9. invoke short-video-production-bible in INTERNAL_STAGE_3 mode with LOCKED_STORY, SELECTED_SHOT_STRUCTURE, AV_PLAN_DRAFT, TOTAL_DURATION, and all available canonical product, prop, character, scene, camera, dialogue, speaking-window, silence, sound, ending-frame, content-boundary, target, and reference fields. Extract compact entity fingerprints, shot-state deltas, boundary states or a deferred shot-boundary ledger, reference needs, critical visual gaps, and explicitly flexible fields; do not duplicate the full story or plan or create a second field vocabulary;
10. resolve blocking production gaps under the existing authority rules. Authorized casting or environment design may be proposed, but remains unapproved until the combined review; a material choice outside authority uses the same concise choice pause;
11. return the compact audiovisual plan and production bible together for one review; stop before long-form prompts.

For a locked blueprint, candidate selection validates the supplied structure rather than replacing it. Build the Beat-to-shot and audience-task coverage maps before product styling or prompt prose. Product installation, feature demonstrations, and result shots may appear only in the canonical Beats that authorize them.

Never copy a reference timeline as a universal structure. FOMO, competitor attacks, provocative claims, ASMR, unboxing, destructive testing, zooms, loops, confessionals, and CTAs are conditional strategies, not required beats.

Phase C — Hand off the approved plan for compilation

Run for VIDEO_PROMPT_SET only after the story and the combined audiovisual plan/production bible are approved, unless the explicit direct-delivery override applies. Build one LOCKED_VIDEO_HANDOFF containing the approved story, SELECTED_SHOT_STRUCTURE, audiovisual plan, APPROVED_PRODUCTION_BIBLE, total duration, target model/platform/country/language, product and character references, product immutable fields and state path, camera locks, exact final dialogue and speaking windows, silence and sound locks, ending frame, and unresolved dependencies. When short-video-visual-prompts is available, use its APPROVED_AV_PLAN_COMPILATION mode and treat its output as the final prompt set. Do not run this skill's compiler again or duplicate the prompt package. If the downstream skill is unavailable, use the engine reference's compact compiler as a disclosed fallback. In either path, preserve the selected audience relation, viewpoint, emphasis, coverage density, cut policy, ending grammar, approved block grouping, actual cuts, compact entity records, state deltas, continuity, exact speech, timing, delivery, and sound without expansion or repetition. Changed variable product lines require local recounting and timing checks before delivery.

If storyboard Panels or reference assets are requested, route the approved plan and bible through optional Stage 4 in references/workflow-routing.md before prompt compilation. This branch prepares visual references only and cannot redesign actual Shots. A reusable project-level Prompt is a later, separately requested deliverable owned by short-video-prompt-architect; do not compile it here.

Validation

Read references/quality-gates.md and run only checks applicable to the selected phase:

• Phase A: causal completeness, provenance, minimal necessary completion, and applicable story-quality gates;
• Phase B: story approval, meaning-preserving duration fit, candidate-specific feasibility, audience-task and source-unit coverage, useful option distinction, selected-structure inheritance, cut justification, shot economy, practical capture style, per-line speaking windows and counts, action/speech feasibility, total time, generation risk, continuity, production-bible completeness/economy, and localization fidelity;
• Phase C: joint plan/bible approval, exact inheritance of final speech/timing and entity/state locks, observable prompt language, continuity, payload economy, and adapter fit.

When a story uses panic or escape behavior, a late reveal, or a product as the reveal object, apply the relevant causal gates:

• PRODUCT_REVEAL_CAUSALITY_GATE
• REACTION_PLAUSIBILITY_GATE
• REVEAL_FAIRNESS_GATE

These are conditional checks, never mandatory plot ingredients.

Output Contract

Pre-Phase output — direction required

For PRODUCT_BRIEF_ONLY, return the direction-stage result produced by short-video-director-thinking. When continued story production was requested, end with one concise selection instruction and set DIRECTION_STATUS: WAITING_FOR_SELECTION; do not append an invented complete story, duration judgment, shot plan, or production bible. If the user authorized automatic selection, state the selected direction once, bind the compact handoff internally, and continue to Phase A without reproducing every rejected option.

After the user selects a direction, do not output a standalone handoff card unless requested. Use it internally to produce the integrated Phase A story, disclose only material assumptions or unresolved dependencies, and request story approval under the normal Phase A contract.

Phase 0 choice pause — either story mode

For an unresolved material conflict or required missing setting, return only the relevant established context, the exact conflict or missing field, and the minimal questions/options under the completion contract. Use 待确认冲突 for incompatible facts and 待选择设定 for missing settings; do not label omission as contradiction. Keep the story DRAFT and stop before rewritten story text or downstream work. After the user answers, apply the contract's local update rules and resume the requested mode. This input-choice pause is separate from approval of a completed story.

Story-completion output with no downstream target

For story_completion with DOWNSTREAM_TARGET: NONE, return:

1. 优化后的完整剧情 — establish character/scene details, supplied dialogue, speech intent or brief draft lines, and meaningful sounds in prose; identify AI lines once as drafts pending audiovisual timing, not a finalized recording script;
2. 关键修复说明 only when a material repair, assumption, addition, or unresolved decision needs disclosure.

Do not append an approval request, storyboard, or video prompt.

Phase A output with an audiovisual-plan or video-prompt target

After Phase 0 decisions are resolved, return only:

1. 优化后的完整剧情 — integrated events, character/scene settings, supplied dialogue, speech intent or brief draft lines, and necessary sounds; not a timed script or separate audio report;
2. 关键修复说明 — only material repairs, proposed assumptions, and deferred dependencies; keep the provenance ledger internal unless requested;
3. 剧情状态 — WAITING_FOR_CONFIRMATION, followed by one concise approval instruction for the integrated version.

Do not append a storyboard or video prompt unless one_shot_override applies.

Phase B output: compact audiovisual plan and production bible

If the preflight returns FIT_PENDING_USER_APPROVAL, return only its compact story-change choices and stop. If SHOT_STRUCTURE_STATUS: OPTIONS_READY, return only two or three compact, viable structure candidates and one recommendation under the selector's schema; do not output multiple complete timed plans or a production bible. After a structure is selected, state the selected structure and its DURATION_FIT_STATUS once, then return 音画分镜草案 using the engine's compact schema: core block range/action chain/capture, actual cuts and their useful purpose, and necessary sound; attach each line's speaking range/speaker/type, exact text, count/budget, and necessary delivery/pauses within its block. State Beat, core Block, actual Shot, Panel when requested, and Generation Segment counts independently rather than treating them as interchangeable. Surface GENERATION_RISK only when HIGH or MODEL_LIMIT_CONFLICT materially affects coverage or packaging. Then return 制作圣经（随音画分镜一并确认） from short-video-production-bible: define stable entity fingerprints once, record per-shot state deltas and necessary boundaries, and include only material reference needs, blocking gaps, or flexible fields. Do not duplicate the full story, option analysis, preflight ledger, or plan. Set AV_PLAN_STATUS: WAITING_FOR_CONFIRMATION and PRODUCTION_BIBLE_STATUS: DRAFT; request one joint approval and stop. A plan-only request ends here without generating final prompts.

Phase C output: locked handoff and video prompt set

After plan approval, pass the following internally to short-video-visual-prompts when available:

LOCKED_STORY
SELECTED_SHOT_STRUCTURE
APPROVED_AV_PLAN
APPROVED_PRODUCTION_BIBLE
TOTAL_DURATION
TARGET_MODEL
TARGET_PLATFORM
TARGET_COUNTRY
VOICE_LANGUAGE
ASPECT_RATIO
PRODUCT_REFERENCES
PRODUCT_IMMUTABLE_FIELDS
PRODUCT_STATE_PATH
CHARACTER_REFERENCES
SCENE_REFERENCES
CAMERA_LOCKS
FINAL_DIALOGUE
SPEAKING_WINDOWS
SILENCE_LOCKS
SOUND_LOCKS
ENDING_FRAME
CONTENT_BOUNDARIES

Then return the downstream compilation result without a duplicate story or plan report:

1. 全局连续性锁定 — only shared settings needed for generation;
2. 模型适配说明 — only when requested or materially relevant;
3. 逐生成段视频提示词 — generation packaging may group approved blocks when the verified model supports it, but may not change actual shots, cuts, event order, dialogue, timing, reveal, or ending;
4. 跨段连续性与拼接 — only when more than one generation segment exists;
5. 负面约束 — only relevant failure risks;
6. 待解决问题 — only unresolved dependencies or conflicts. Keep passing checks and the already reviewed plan internal.

If the downstream skill is unavailable, return the engine fallback schema and label it FALLBACK_COMPILER.

If total duration is missing, ask once or return only an untimed beat plan labeled TIMING_NOT_ASSIGNED when the user requests progress without it.

If FIT_PENDING_USER_APPROVAL or DURATION_CONFLICT is found, return the duration reference's concise analysis and smallest alternatives; stop before an invalid plan or final prompts. Continue after the user selects a repair or changes duration. A changed frozen beat/meaning requires story review; revised final lines or windows require plan review only, subject to approved variable scope.

Use plain text unless the user requests another format.

Adjacent Skill Handoff

This skill is the canonical public entry for a product brief, selected direction, submitted outline, or story. It routes product-only inputs through optional Stage 0 and owns Stages 1–3 in references/workflow-routing.md. short-video-production-bible is invoked inside Stage 3 and does not create a separate mandatory user step.

After the joint audiovisual-plan/production-bible approval, use optional short-video-storyboard-prompts only when the user needs Panels, mother images, or reference assets. It inherits the selected structure and actual Shots and may not redesign them.

Route final one-off prompt compilation to short-video-visual-prompts in APPROVED_AV_PLAN_COMPILATION mode using the locked handoff above. Its configurable defaults may not override this skill's approved story, selected shot structure, entity fingerprints, scene layout, state deltas, dialogue, timing, silence, editorial Shots, reveal, product-state, camera, boundary, or ending decisions.

Only after the final one-off video-prompt set exists, if the user asks to build, revise, or audit a reusable project-level short-video Prompt, route that work to short-video-prompt-architect; do not perform the abstraction inside this skill.

When a user supplies an approved reusable specification from that skill together with a concrete product or story, treat the specification as upstream constraints and decision criteria. Continue to enforce this skill's story-state and approval gates. Return observed framework defects as proposed upstream changes; do not silently rewrite the reusable master prompt during one-off production.

Non-negotiable Boundaries

• Do not run direction ideation for an input that passes the usable-story test unless the user explicitly requests new directions. A weak or incomplete synopsis is still a story-completion problem, not permission to replace its premise.
• Do not treat a recommended or selected direction as a completed or approved story. Story integration and story approval remain separate.
• Do not pass rejected direction options or verbose upstream analysis downstream; preserve only the selected direction, facts, locks, boundaries, and unresolved dependencies.
• Story completion means closing necessary gaps, not explaining every inference. Preserve useful uncertainty and negative space.
• Observable absence can be sufficient evidence: when an expected change repeatedly fails to occur, characters may cautiously update their behavior without receiving a new message, clue, or explanation.
• Do not pre-solve a hard reveal. Add explicit foreshadowing only when the chosen reveal style or user request requires it.
• Do not force one product, character, country, language, setting, duration, platform, viewpoint, shot count, model, CTA, FOMO device, unboxing beat, or before/after structure without an applicability rule.
• Do not force fixed time ranges; timing must be derived and must sum exactly to the requested duration.
• Do not declare an approved story infeasible before testing non-material screen compression. Apply meaning-preserving timing, overlap, action-chain, transition-ellipsis, and editing operations without a redundant approval; never disguise a material deletion or reinterpretation as timing. Pause with FIT_PENDING_USER_APPROVAL when an unapproved story change is required, and use DURATION_CONFLICT only after credible non-material packing is exhausted or maintained locks make the target impossible.
• Do not require spoken audio in every segment. Silence and native sound are valid decisions.
• Do not treat absent source dialogue as approved silence. Complete SPEECH_DECISION before story review, and let intentional silence be a visible creative decision.
• Do not use polished marketing copy, generic explanatory dialogue, or slogan-shaped reactions as a substitute for character speech. Product information enters speech only when the approved character purpose and supplied facts require it.
• Do not require sound effects to overpower speech; establish an audio priority appropriate to the beat.
• Visible proof establishes only what the visible action supports.
• Optional upstream information may improve the result but must not become a hidden dependency.
• Character and scene fields are selected by their story function, not a fixed persona checklist; use the completion contract's role and utility rules.
• Regional scene design belongs before story approval; spoken localization belongs to the optional adapter. Neither implies the other.
• Preserve user-selected dramatic intensity, including police, weapons, threats, panic, coercion, injury, crime coding, or other high-intensity material. Evaluate it for narrative purpose, causality, physical staging, timing, continuity, and model execution; do not lower or replace it for platform-policy, ad-review, brand-safety, or audience-sensitivity reasons unless the user explicitly requests that review.
• Model limits belong in a verified, replaceable adapter.
• Revise the smallest affected beat, scene, shot, segment, or conditional decision before rebuilding the whole result.