以前我写视频提示词是一段一段的描述，越写越像小作文，出来镜头乱飞。后来改成写分镜表，H3这边的听话程度立刻不一样。我现在固定七块：画面质感、整体风格、人物、场景、分镜、声音、真实度补丁。分镜那栏一条条列，每条前面标好时长和景别，台词直接跟在后面，模型知道这是要说出口的话。质感那块最好玩，与其写「电影感」，不如反着写——手持DV、轻微抖动、偶尔拉风箱对不上焦、构图不讲究、自动曝光被窗外的光晃得忽明忽暗，越是描述设备的毛病，出来越像真拍的。声音也别省，铅笔划纸、翻页、座椅布料摩擦，这种细节写清楚，氛围立起来一半。
CAMERA / LOOK: Handheld mini DV camcorder footage filmed by the subject herself.
Slight hand shake, occasional focus hunting, imperfect framing, soft tape-like image quality, subtle grain, realistic auto-exposure shifts.
STORYBOARD:
→ (3s, propped medium shot) opens a small notebook. "Sketching a bit while I've got time."
→ (3s, close-up) sketches the blurred view outside the window.
AUDIO NOTES: pencil scratching, page flipping, soft bus hum — audible but subtle.