short-drama-assets

Short-Drama Asset Breakdown

Turn script text into reusable, sustainable visual facts, written uniformly into 剧集/<EP>/视觉设定.md. Do not turn every appearance into a new asset, and do not take over writing script semantics, camera blocking, or generation prompts.

Quick Start

用 $short-drama-assets 从 EP001 剧本拆完人物、造型、地点、道具和本集状态

Entry Points and Boundaries

An existing script can be broken down into assets directly. First distinguish:

• Identity: if it is swapped for another, it is no longer the same person, place, or thing;
• Variant: identity stays the same, but clothing, injuries, time period, weather, open/closed state, or held state changes;
• Shot transient: pose, gaze, left/right hand, blocking, and camera angle are owned by the storyboard;
• Story semantics: knowledge, goals, relationships, and emotions are owned by writing/development; only reference visible consequences.

Workflow

1. Read only the script in the current scope and the already-decided visual direction; when the visual direction does not yet have an era anchor, first set the era according to the source material's facts.
2. Scene by scene, find entities, visible descriptions, state changes, and ambiguous references; do not create assets yet.
3. For each item, decide create new, reuse, add variant, or mark undecided; do not guess at ambiguous references.
4. Write the minimum recognizable setting: uniformly use 识别锚点, this episode's state, and the reason for change; body-object-spatial relationships that will persistently constrain later actions should be written into states with a scope of effect, not skipped as a one-off pose.
5. Check looks/injuries, held items/ownership, prop states, location time period/weather/light state, and cross-episode outgoing.
6. Write one continuity lock for visible facts that appear across shots and whose inconsistency the audience can notice; for the rest, keeping recognition anchors is enough.
7. When the project prompt language is not Chinese, write one line of 画面代称：<正文里的拼写>for each character entry; if the main text never names it, write 画面代称：无. Locations and props as needed.
8. A full-episode request automatically handles the complete episode, and at the end reports key identity decisions and undecided references once.

Content Requirements

• For each item, first write what does not change so that it remains the same asset, then write this episode's changes.
• Recognition anchors must be visible, generatable, and comparable; do not replace facts with vague quality words.
• Clothing, injuries, wet stains, prop open/closed state, and light state are usually variants, not new identities.
• Do not copy plot summaries; keep only the facts that downstream generation, storyboarding, or continuity truly need.
• Text, logos, and sensitive information follow project policy; do not arbitrarily invent brands or private information.
• For multi-episode states, write clearly the effective and continuous scope; leave in-shot transients to the storyboard.
• Confirmed relationships that persistently constrain the action space (bodily support, fixation, load-bearing, occupancy, or contact with a functional object) are not ordinary shot poses: independently recognizable objects are still built as assets, while the relationship is written into the state and handed to the storyboard for shot-by-shot projection; it ends only when the script explicitly changes it.
• Continuity locks lock only visible facts that do not change with the plot; the locked surface is written in the prompt language as a minimal noun phrase, and the parts that will change are left to the state.
• "Screen alias" maps the names in the prompt body back to this entry: the storyboard's mechanical verification relies on it to determine who the person named in the keyframe is. The entry's own Chinese name is always counted as an alias, so the Chinese body usually does not need this line; when the prompt body language is not Chinese, a character entry must give a conclusion—write the spelling used in the body, or write 画面代称：无when the body never names it. The same name set uses only one spelling, with consistent capitalization (AST-10 the same requirement holds across languages); separate multiple spellings with 、 . When a name is homographic with a common word (道具「手机」 clashes with "phone store"), also write 画面代称：无, so that verification does not look for it by name.

On-demand knowledge

By default, only read this SKILL, the current script, and the visual direction. When encountering a corresponding issue, open only one:

• Stage boundaries and rule grading:Stage contract
• Setting the era, writing era anchors, or wanting to make places, equipment, or clothing look old:Era anchors
• Finding actual asset appearance evidence from the script:Appearance extraction
• New identity, reuse, variant, or undecided:Identity and variants
• Character recognition and styling layers:Characters and styling
• Location geography, views, and fixed anchors:Locations and views
• Functional props, ownership, state changes, or readable text and signage on screen:Props and states
• Identity and authorization boundaries of voice references:Voice direction
• Cross-scene, cross-shot, cross-episode states:Continuity changes
• Cross-shot clothing/prop color and material drift:Cross-shot consistency lock
• Do an asset omission check before completion:Asset review checklist

Completion

All clear appearances have been reused, newly created, variant-ized, or marked as genuinely undecided; the visual facts and state changes that need to be maintained have been written into the single visual settings document, and it is complete. Image prompts, storyboards, review, and production do not start automatically. Once the five creative documents are complete, you may switch to $short-drama to perform a mechanical check of the cross-document structure; content quality is still reviewed by the creator.

Installation and maintenance

Run only when installing, upgrading, or troubleshooting python3 scripts/selftest.py。