Short Video Visual Prompts

Short Video Visual Prompts

Translate settled creative intent into instructions a video model can see or hear. In the canonical submitted-story workflow, this is Stage 5 after joint audiovisual-plan and production-bible approval and after any requested reference assets are prepared. Preserve upstream authority. This skill is a production compiler and repair tool, not a second story writer.

Authority Order

Apply inputs in this order:

1. current explicit user instructions;
2. APPROVED_AV_PLAN, APPROVED_PRODUCTION_BIBLE, LOCKED_STORY, exact dialogue, approved silence, and other named locks;
3. supplied product, character, scene, audio, and reference assets;
4. verified target-model capabilities and syntax;
5. configurable production defaults;
6. examples and templates.

Lower levels may fill open execution details but may never replace higher-level decisions. When two locks conflict, return LOCK_CONFLICT with the exact fields. Do not choose whichever is easier to generate.

Modes

Infer the narrowest mode; do not ask the user to choose from a menu.

• APPROVED_AV_PLAN_COMPILATION: an approved timed audiovisual plan or locked handoff package exists. Read references/approved-av-plan-compilation.md. Compile only; do not redesign.
• VISUAL_PLAN_FROM_STORY: direct-invocation fallback only when the user supplies a settled story, explicitly asks this skill to design shots or prompts, and no meaningful duration/structure approval choice is pending. Read references/visual-plan-from-story.md. If duration fit is uncertain or materially different shot structures should be offered, route planning to story-to-reusable-prompt instead of bypassing its review gate.
• SEGMENT_REPAIR: the user identifies a failed or weak generated segment, prompt, action, product rendering, voice, or continuity boundary. Read references/segment-repair.md.
• STITCHING_ONLY: the user already has generated clips and asks only for ordering, head-tail alignment, audio continuity, or edit notes. Read references/stitching-only.md.

Mode priority is approved-plan compilation, explicit local repair, stitching-only, then visual planning. A request to “turn this approved plan into prompts” is compilation even if the skill also knows how to design shots.

Input Contract

Capture only relevant fields:

MODE
SOURCE_AUTHORITY
LOCKED_STORY
SELECTED_SHOT_STRUCTURE
APPROVED_AV_PLAN
APPROVED_PRODUCTION_BIBLE
TOTAL_DURATION
TARGET_MODEL
TARGET_PLATFORM
TARGET_COUNTRY
VOICE_LANGUAGE
ASPECT_RATIO
VISUAL_PROMPT_LANGUAGE
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

Do not turn an optional field into a hidden dependency. Never place unresolved markers such as 【待确认】, 【待证据确认】, TBD, TODO, UNKNOWN, or [placeholder] inside a generator-facing prompt.

Classify a missing field before compilation:

• If it is an open execution detail that can be omitted or safely resolved by an existing default, continue without mentioning the omission inside the prompt.
• If it materially controls a visible action, product state, exact factual line, safety-sensitive operation, required result, or generation packaging, ask the user one concise grouped clarification and pause compilation. Do not emit a partial “direct copy” prompt.
• If the user explicitly asks for a placeholder draft, keep unresolved markers only in a separate non-executable planning or 待解决问题 section. Do not place them in 直接复制用综合提示词, and do not label that draft as executable.

An approved story result is creative authority, not automatically an unresolved dependency. Compile its observable result as approved unless the user explicitly requests real-product verification, claim substantiation, regulatory review, or exact operating instructions. Lack of external proof does not authorize adding caveats to the video prompt. It also does not authorize inventing dosage, temperature, reaction time, chemical mechanism, safety procedure, or stronger effects than the approved result.

When APPROVED_PRODUCTION_BIBLE is supplied, use its canonical entity fingerprints, scene layout, shot-state deltas, boundary records, and reference roles as authoritative production facts. Its absence is not automatically blocking for a fully concrete legacy plan; however, if compilation would require first-time invention of material casting, environment, product geometry, action state, or continuity, return PRODUCTION_BIBLE_REQUIRED and route that normalization to short-video-production-bible before generating prompts.

Shared Rules

• Describe observable or audible execution: starting state, action path, ending state, subject position, product state, camera access, light, sound, and boundary continuity.
• Preserve event order, reveal timing, character knowledge, approved intensity, selected audience relation, viewpoint strategy, priority meaning, coverage density, cut policy, ending grammar, shot count, actual cuts, continuous-take decisions, exact dialogue, speaking windows, pauses, and intentional silence.
• Do not force early product appearance, product dominance, a demonstration, proof shot, hand close-up, reaction, CTA, FOMO, before/after structure, captions, background music, hard cuts, handheld movement, or a fixed shot duration.
• Product appearance and claims come only from supplied facts and references. Do not invent logos, text, price, inventory, certification, endorsement, superiority, or effects.
• Distinguish an approved observable story outcome from an invented mechanism. For example, an approved drain-clearing result may be shown as the locked outcome; do not add unsupported dosage, water temperature, “instant reaction,” “emulsification,” “decomposition,” or complete restoration language.
• Keep character, product, prop, space, light, camera direction, action vector, emotion, knowledge, dialogue, and ending frames continuous across generation boundaries.
• Resolve each approved Block/Shot to its production-bible entity IDs and local state deltas before prompt prose. Do not replace an ID with a newly invented character, environment, product, or prop description.
• Speech follows the approved speaker, addressee, language, meaning, timing, delivery, and mouth visibility. Compile each final line once. Do not turn dialogue into voiceover or marketing copy.
• A user-supplied target country determines the language of speech that exists unless the user explicitly supplies VOICE_LANGUAGE. It does not create speech or change approved silence.
• Use the fewest generation segments supported by the approved plan and verified target model. A story beat, shot, core block, and model call are not automatically the same unit.
• For each independent generator call, select only the entity fingerprints, reference assets, entry state, action path, exit state, speech/sound, and boundaries used by that call. Never paste the complete production bible into every call. Repeat a compact identity anchor once only when that call must stand alone.
• Treat model duration, reference limits, audio support, resolution, and syntax as replaceable adapter facts. Verify them when model-specific output is requested; if verification is unavailable, state the assumption and keep the core prompt model-neutral.
• Explain in Chinese by default. Keep final prompts in the user's requested language; retain spoken lines in their approved language.

Clarification and Prompt-Cleanliness Gate

Before writing any generator-facing payload:

1. List only the unknowns that truly affect what the model must render or say.
2. Resolve non-critical open fields by omission or approved defaults.
3. If any critical unknown remains, ask for it in a concise user-facing question and stop before producing prompts.
4. If no critical unknown remains, compile the prompts with no internal workflow labels or unresolved annotations.

Every section labeled 直接复制用综合提示词, GENERATION_PROMPT, or equivalent must be clean input for a video model. It must contain only renderable audiovisual instructions and supported model syntax. It must not contain:

• evidence requests, review comments, approval notes, confidence labels, or compliance commentary;
• bracketed uncertainty markers or instructions for a human to fill later;
• UNVERIFIED, PLACEHOLDER_DRAFT, REQUIRED_INPUT, or similar workflow state tokens;
• production notes that would risk becoming generated captions, labels, signs, narration, or spoken dialogue.

Keep internal status and unresolved issues outside the generator payload. When a missing answer changes that payload, clarification is required first; a warning beside the prompt is not an acceptable substitute.

Default Decisions Only for Open Fields

When no approved plan or user choice controls an open production field:

• prefer practical framing and physically continuous action;
• group compatible beats rather than cutting on every reaction or speaker turn;
• use hard cuts only when a cut is actually needed;
• use native environment and action sound before optional music;
• omit captions unless requested;
• choose aspect ratio from the target platform when reliably known, otherwise leave it configurable;
• choose product-entry timing from the story function, not a universal early-product rule.

These are fallback decisions, never locks.

Output by Mode

Approved plan compilation

Return:

1. 全局连续性锁定 — only canonical facts shared by the delivered generation segments;
2. 模型适配说明 only when a target model is requested or an assumption matters;
3. 逐生成段视频提示词, preserving approved block/shot boundaries and exact speech;
4. 跨段连续性与拼接 only when there is more than one generation segment;
5. 负面约束 limited to relevant failure risks;
6. 待解决问题 only for non-blocking dependencies that do not change the supplied generator payload. Ask about blocking unknowns before compilation.

Visual plan from story

Return the requested setup, timed shot or segment plan, prompts, continuity, sound/voice, stitching, and relevant negative constraints. Do not output every section when the user asks for only one part.

Segment repair

Return the diagnosed failure, smallest changed prompt or boundary, preserved locks, and local revalidation result. Do not rewrite unaffected segments.

Stitching only

Return clip order, boundary-by-boundary visual and action alignment, audio/dialogue handling, and edit instructions. Do not generate replacement story beats or prompts unless explicitly requested.

Validation

Before delivery, check only applicable gates:

• AUTHORITY_GATE: no lower-priority default replaces an upstream lock.
• LOCK_COVERAGE_GATE: every approved event, shot, line, window, silence, reveal, and ending remains present once and in order.
• STRUCTURE_INHERITANCE_GATE: SELECTED_SHOT_STRUCTURE remains authoritative; compilation and model packaging do not silently change viewer relation, viewpoint, emphasis, density, editorial cuts, reveal strategy, or ending grammar.
• SHOT_COUNT_IDENTITY_GATE: editorial Shots remain distinct from core Blocks, storyboard Panels, and Generation Segments; model calls may package but not redesign them.
• PRODUCTION_BIBLE_GATE: supplied entity fingerprints, scene layout, state deltas, reference roles, and boundaries are inherited without recasting or first-time production design.
• PAYLOAD_ECONOMY_GATE: each generator call includes only its referenced entities and local states; the full bible is not duplicated across calls.
• OBSERVABILITY_GATE: prompt prose describes visible or audible execution.
• MODEL_ADAPTER_GATE: current model facts are verified or clearly labeled assumptions.
• DURATION_GATE: generated segments cover the approved total with no gaps or overlaps and fit the model adapter.
• CONTINUITY_GATE: boundary states match or show a motivated transition.
• PRODUCT_GATE: product state and immutable appearance remain consistent.
• AUDIO_VOICE_GATE: speech type, exact line, language, timing, delivery, and mouth state are inherited correctly.
• PROMPT_CLEANLINESS_GATE: every generator-facing payload contains no placeholders, evidence requests, workflow status, human-facing notes, or other text that could be rendered into the video.
• CLARIFICATION_GATE: any unresolved choice that would materially change visible action, speech, safety-sensitive operation, result, or packaging was asked before prompt generation.
• LOCAL_REPAIR_GATE: repair mode changes only the failed scope.

Use references/regression-cases.md when materially revising this skill. Test routing and invariants, not exact prose.

Adjacent Handoffs

• Use story-to-reusable-prompt as the canonical main entry for rough-story repair, character/scene integration, speech intent, duration fit, shot-structure selection, story approval, and joint audiovisual-plan approval. Once it supplies APPROVED_AV_PLAN and APPROVED_PRODUCTION_BIBLE, enter Stage 5 compilation here.
• Use short-video-production-bible before plan approval to normalize production-critical character, scene, product, prop, action-state, and boundary locks. Consume its approved output here rather than recreating those decisions.
• Use short-video-storyboard-prompts as optional Stage 4 for storyboard pages, Panel layouts, mother images, and reference-image planning. When those assets will guide generation, consume them here after they are prepared.
• Use short-video-prompt-architect only after this skill has produced the final one-off prompt set, when the user wants a reusable project-level Prompt.
• Do not generate media, publish, or operate external video tools unless the user separately requests and authorizes that action.