Principles
- Control comes from blocking/light/physics/counts, not masterpieces.
- Write only this shot; delete scene numbers, expired
@, characters not yet on screen. - Present tense, short sentences, positive phrasing.
- With reference: minimal anchor +
100% matches reference。
Inputs from: script Brief + character anchors/voice/performance-scene changes + storyboard blocking;GEO/Style constants are not reinvented, only pasted.
Video Skeleton
SCENE CONTEXT EXACT N CHARACTERS — NO DUPLICATES
ACTIVE REFS 本镜标签 + 职责（角色/场地）
LOCATION MAP/GEO 场地参考不继承构图
FIRST FRAME 占位、视线
FORMAT 单 take 或逐刀 multi-shot
OPTICS / CAMERA 可见光学；唯一主运动
ACTION TIMING 每拍≤3 句
PHYSICS / LIGHTING
AUDIO 声线原样 + 引号台词；SFX；无配乐字幕
CHARACTER ACTING 欲望/阻碍/可见变化
STYLE / QUALITY Style Prefix 原文
POSITIVE LOCKS 人数道具伤位机位轴 — 写「画里有什么」
long can be 3k–4k words;do not overload a single take with mutually exclusive elements.
STYLE and this shot's LIGHTING: project Prefix defines the medium base; this shot's light may specify concrete top light/backlight, do not let the two strangle each other (on conflict, this shot's LIGHTING prevails and revise the Prefix or delete the conflicting sentence).
Blank project Style template:templates/style-prefix.md。
Write the first four steps
deconstruct this shot → diagnose (empty first frame/misplaced positions/gaze/floating/timestamps) → build (space→light-lens→action→performance/voice) → deliver only a paste-ready draft.
5 checks before sending
- Headcount lock —
EXACT Nmatches the cast list, no ghost people - First-frame placement — people/objects that should be in frame are already in place in the first frame (unless an empty shot is designed)
- GEO / crossing the line — frame side and camera axis are clear, no lost scene lock
- Single camera move — only one primary motion (or an explicit shot-by-shot breakdown)
- Dialogue only in AUDIO — no spoken words in action blocks; fixed voice lines + quoted lines; no music or subtitles
Optics / light / physics
| Content | Optical tendency |
|---|---|
| Environment/close-up | wider: environment fills the frame, foreground expansion, no fisheye |
| Portrait/peeping | longer: distant subject shot close, background color blocks, foreground blur allowed |
| Detail | separate insert |
Write visible results (sense of field of view + distance), do not pile up brand focal lengths. Handheld = breathing micro-corrections. One shot, one primary motion.
Light: key source, direction,whether the camera is on the light side or shadow side, exposure priority.
Physics: mass, contact, clothing/hair lag, liquid, weapon weight; no floating.
Performance and dialogue (written layer)
Details in character-design; here only put on paper:
- visible behavior, do not write empty sad/angry
- eyes before head; reactions before the end of the line
AUDIO only:
[Voice 固定句]
Line, nothing else: "..."
Others silent. No music. No subs.
Separate action blocks and dialogue blocks.
Count lock (on-screen evidence)
Exactly 3 people. Exactly ONE crystal on right arm wrist→shoulder.
FIVE wrecked dummies, never intact. Scale: guardian 5× human height…
Image prompt (asset-oriented · short)
80–150 word natural paragraph; positive phrasing; 60/30/10 comes from the scene.
Character-sheet wording: film character sheet / studio photographs (avoid illustration trigger words).
Image editing:one change at a time + PRESERVE list. When changing camera positions, specify left-right mirroring of objects.
Iteration
| change one line at a time | leave everything else untouched |
| log | Version/Changes/Verdict |
| 10–15 rounds fail | Concise mirror, not word-adding |
| Diagnose | Empty shot→first frame; reposition→GEO; plastic→styling/light; fake acting→performance section |
Build your own banned-word list. Write fewer age numbers. Don't look at the camera unless designed.
Short anchor
[景别] of @A and @B in @Loc. [动词]. [一句光].
@A: "..."
Compare with sample text
examples/training-hall-12s.md — Training hall 12s English (generated) / Chinese (comparison). For new shots, copy the structure then change the content.
Final-cut pacing and cleanup → edit-assembly。