physical-action-videolisted

# Physical-Action Video (manipulation-beat craft) This is the **video-craft** layer for the hardest beats: a character physically interacting with an object — filling, opening, pouring, applying, twisting, gripping-and-releasing, writing, cutting, lifting. Frame analysis CANNOT validate these (each frame in isolation is plausible; the error is in the *motion*). The defense is BOTH (a) explicit prompt decomposition AND (b) first-last-frame generation + a model-escalation ladder + post-generation video-understanding verification. The Storyboard owns the *workflow* (which card, when to spend); `ai-asset-generation` owns the *call + save mechanics*; `realistic-image-generation` owns the *keyframe images*. This skill owns *how to make a manipulation beat not fall apart.* Apply it whenever a beat is a physical manipulation (`physicalActionVerification` beats in `ugc-product-video`). ## Stage 4.5 video-understanding question bank (cross-reference) When the beat is flagged `physicalActionVerification: true`, the agent's Stage 4.5 video-understanding pass uses a TWO-PART question set: 1. **Universal physical-plausibility checklist** (6 fixed questions, always asked — covers liquid source, object permanence, gravity, grip stability, spatial geometry, on-screen text). 2. **Beat-specific questions** (2–4 generated from the beat's decomposition sub-steps). See `ugc-product-video` Stage 4.5 for the canonical question text. The decomposition you write here IS the source-of-truth for th