What problem does it solve? Video generation models receive only a single prose prompt per shot, so multi-step actions, camera moves, and timing are often ignored or improvised. This Skill pins down each shot's motion second-by-second before any paid video generation, using a cheap sketch board plus a zero-cost segmented text timeline. ## Core Features & Use Cases - Per-second beat planning: Authors write shots[].sketch.beats (2-3 seconds per beat) that the engine compiles into a segmented timeline injected into the video prompt, with model-specific timestamp formats. - Sketch board generation: sketch gen renders a pencil-sketch grid (one panel per beat) with color-coded arrows for subject motion, camera moves, framing, lighting, and sound, costing about 0.3 CNY per board. - Mutual-exclusion arbitration: Each shot picks exactly one motion guide (sketch, 3D previz, or depth-control video) via sketch use --guide, preventing conflicting motion instructions. - Use Case: Before generating an 8-second fight shot with Seedance, write 3-4 beats with explicit timing, generate one board to verify staging, then run gen-video --dry-run to audit the timeline prompt before spending on real generation. ## Quick Start Ask the agent to write per-second beats for the shots in a chapter and generate a sketch previz board for the first shot to verify the staging before video generation.
