界面已本地化;技能介绍与说明文档按所选语言由 AI 自动翻译,代码、命令、路径和专业术语保留原文,以原作者文档为准。
探索技能
设计与创意 · SKILL.md

shot-design

shot-design

【设计与创意】Design motivated shots from a screenplay scene and director book, giving every shot a purpose, framing, camera behavior, action, continuity state, duration, and sound plan.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准

技能说明

GitHub

name: shot-design description: Design motivated shots from a screenplay scene and director book, giving every shot a purpose, framing, camera behavior, action, continuity state, duration, and sound plan. disable-model-invocation: true author: Alan Guice (Badgids) license: Apache-2.0

Shot Design

Workflow

  1. Read core contract, VISIBLE_DIALOGUE_SYNC.md, dramaturgy rules, film grammar, prompt packet schema, screenplay scene, line manifest, breakdown, director book, continuity, production capabilities when present, performance blocking when present, and approved reference manifest.
  2. Assign stable SHOT-### IDs within the project sequence.
  3. Design the minimum coverage needed for the intended cut.
  4. Write one model-neutral JSON object per shot to 04_generation/shot_briefs.jsonl.
  5. Each shot must include line_ids for the source production units it covers when a line manifest exists, plus purpose, dramatic job, beat, scene ID, duration, subjects, location, continuity, framing, composition, camera, movement reason, eye trace, action, environmental pressure, micro-action, anchor, lighting, dialogue, ambience, music, SFX, references, cut intent, and constraints.
  6. Resolve applicable visual-bible capture behavior into capture_behavior without inventing camera hardware.
  7. When the viewer must see a character speak an exact screenplay line, add the model-neutral lip_sync record with exact LINE-###, speaker, visibility, cut policy, and measured timing when available. Do not force lip sync onto off-screen dialogue. When approved audio exists, preserve its MEDIA-###, SHA-256 authority, and start time and run dialogue timing preflight before expensive generation.
  8. Add end_frame only when chained generation, match on action, last-frame conditioning, or editorial continuity needs an explicit final state. When motion state also matters, add a temporal-continuity handoff that uses the approved previous shot and a visual-only tail; do not import prior-shot audio.
  9. Add optional frame_regions only when multi-subject screen placement needs explicit control or the downstream route supports regional conditioning. Use semantic zones unless normalized bounds were deliberately planned.
  10. Check screen direction, eyelines, eye trace, and geography across adjacent shots.
  11. If the requested camera/action combination is unavailable or conditional in production_capabilities.json, redesign, split, or mark the blocker rather than pretending it is executable.
  12. Delete shots that do not change a relationship or emotion, advance action or information, increase pressure, or satisfy an explicit practical coverage need.

Done

Every shot earns its place in the edit and can be generated independently from its saved brief.

我的收藏