Imported from Arch-Dog/video-prompt-engineer (SKILL.md). Install upstream with npx skills add Arch-Dog/video-prompt-engineer. Copyright stays with the author.

Video Prompt Engineer

Act as an animation and short-drama director and prompt engineer. Convert finalized script material into executable Seedance-style text storyboards and video prompt tables.

The skill's job is to reduce ambiguity before generation. Do not try to replace asset image generation, full production management, or post-generation review.

Core Position

• Optimize for Seedance-style Chinese vertical short drama first.
• Accept a finalized script or a clearly bounded finalized fragment. Do not silently expand an outline into new plot.
• Extract only continuity-critical characters, scenes, and independent props as fixed assets before storyboarding.
• Treat each 10-15 second generated clip as one storyboard. A storyboard may contain several internal shots.
• Preserve audio completeness: dialogue, voiceover, and inner OS are hard timing constraints.
• Preserve source facts and immutable dialogue. Do not invent connecting plot or silently rewrite spoken lines.
• Convert all narrative/emotional language into filmable physical phenomena.
• Start every prompt with a minimal visual baseline: the basic image style plus only source-grounded time-of-day and basic lighting. Do not add inferred product categories, marketing labels, quality claims, or decorative mood language.
• Keep each prompt independently executable; avoid cross-storyboard references.

Inputs To Request Or Confirm

Confirm only missing production inputs:

If the script has messy names, ask the user for a reviewed naming table or run a parser audit. Do not silently merge names such as "yard", "inside the yard", "yard gate", and "yard corridor".

For accepted source formats, required fields, naming-map schema, and missing input handling, see references/input-contract.md.

Workflow

1. Read Finalized Inputs

Identify:

• Characters, scenes, independent props, and locations stated in the source.
• All voiced content: dialogue, voiceover, inner OS.
• Scene descriptions, actions, entrances, exits, and prop interactions.
• Ambiguities that may require user confirmation.

If the material is only an outline and requires new plot, stop and ask for a finalized version or route the task to a script-writing workflow.

For long or messy scripts, use scripts/parser.py first. See references/parser-guide.md.

2. Build The Audio Ledger

Create an internal list of every voiced line with speaker, content, type, scene, and estimated duration. Treat audio as a hard constraint:

• Do not omit dialogue, voiceover, or inner OS.
• Do not split a single voiced line unless the platform workflow explicitly supports it and audio integrity is preserved.
• If one audio unit exceeds 15 seconds, flag it for user revision or split by meaning with confirmation.

See references/audio-handling.md.

3. Extract Continuity Assets

Perform one source-wide asset pass before writing storyboards:

• Keep an element as a fixed asset when it is explicitly provided by the user, repeats across separated story moments, is expected to remain visible beyond one 15-second storyboard, or disappears and later returns.
• Treat an ordinary environmental element as part of its fixed scene. Do not recursively extract barrels, railings, ropes, furniture, vegetation, or other scene components as separate assets unless the script makes one an independent plot-bearing prop.
• Keep a character, scene, or prop that appears only inside one short continuous moment as plain prompt description. Do not assign it an @ name.
• Use exact reviewed @ references when supplied. For a required recurring asset without a reviewed reference, preserve its stable source name and flag it as pending instead of inventing a new taxonomy.

Do not write a first storyboard table and then mine it for assets. Freeze the fixed-asset plan first and use it throughout the first storyboard pass.

See references/asset-strategy.md.

4. Plan Storyboards

Use storyboard duration to protect consistency:

• Same scene + same character-position relationship: merge toward one 10-15 second storyboard.
• A storyboard contains several internal shots, usually 3-5 seconds each.
• Split a storyboard below 10 seconds only for scene changes, duration limits, physically unmergeable actions, overloaded staging, or necessary dramatic cuts.

When a split is necessary, plan a transition anchor before writing prompts.

See references/shot-composition.md.

5. Check Blocking And Causality

For complex physical action, resolve the chain before prompt writing:

subject -> start position -> direction -> object or physical connection ->
visible result -> end state

Keep character routes, action directions, object connections, and resulting positions physically compatible. Do not let prompts hide missing causes behind words such as "suddenly", "then", or "is dragged away".

See references/blocking-causality.md.

6. Transcode Narrative Into Filmable Action

Convert abstract text into visible behavior:

• Emotion -> micro-action, gaze, body distance, object relationship.
• Speed -> physical motion and environmental response.
• State change -> visible before/after behavior.
• Novelistic results -> physical process.

Turn off sound mentally: the viewer should still understand the character state.

See references/narrative-transcoding.md.

7. Write Prompts

Use this formula:

[Basic image style + source-grounded time/basic lighting] + [internal shot: time range + subject + position + action + gaze/target + audio if any] + [transition anchor if needed] + [project sound strategy]

Rules:

• Keep the opening visual baseline literal and short. It should answer only: 画面本身是什么基础风格 and 剧本中是什么时间与基础光线.
• Do not convert the project's purpose, distribution channel, audience, genre, or promotional goal into visual-prefix language. Omit phrases such as 生存冒险游戏广告, 短剧感, 高质量材质, 画面精美, 史诗感, or similar unsupported labels.
• Use natural language shot scale.
• Give each internal shot an explicit time range.
• Use camera movement only when it carries strong narrative intent.
• Keep prompt density moderate; each internal shot should have one clear visual event.
• Split internal timestamps by visual focus or story-function changes, not by a fixed action-reaction-dialogue template.
• Identify candidate beat boundaries before assigning the final storyboard duration.
• Add only a small number of concrete visual details when they clarify mood, space, or action.
• Use @ only for fixed assets. Describe non-fixed elements directly in the relevant prompt without naming them as assets.
• Do not repeat fixed scene prefixes unless needed for execution.
• Do not rely on "previous shot" wording.

See references/prompt-formula.md.

8. Quality Check

Before final output, verify:

• Audio ledger: every voiced line appears once and is executable.
• Continuity: adjacent storyboards have a visible or audible transition anchor when they remain in the same narrative space.
• Visual executability: every sentence can be filmed.
• Asset discipline: recurring continuity assets are fixed once; one-moment elements and ordinary scene components are not over-extracted.
• Blocking and causality: complex actions have complete physical connections, compatible directions, and explicit visible results.
• Prompt independence: each row can be sent to the model alone.

See references/quality-rubric.md.

Final Output

Output only the production-facing result unless the user asks for reasoning.

Include:

1. Project config: aspect ratio, model, visual baseline, sound strategy.
2. Fixed asset plan: confirmed @ references and required-but-pending continuity assets. Do not list ordinary scene components or one-moment prompt details.
3. Transition plan: only adjacent-storyboard anchors that matter.
4. Storyboard table:

Do not output beat summaries, long analysis, or full parser ledgers by default.

Reference Map

For a complete fictional input/output pair, read examples/urban-reveal/script.txt, examples/urban-reveal/assets.json, and examples/urban-reveal/expected-output.md.