What problem does it solve? LLMs can read images but cannot watch videos, making it impossible to analyze screen recordings, user complaint clips, or interaction bug footage. This Skill uses pure code (frame-differencing) to turn a video into timestamped keyframes plus a motion timeline, so the model can tell you what happened at second X without wasting tokens on still segments. ## Core Features & Use Cases - Zero-token frame selection: Frame-differencing runs on CPU via OpenCV to detect which seconds are moving, automatically folding away still segments before the model sees anything. - Coarse-to-fine workflow: Use scan for a motion timeline of the whole clip, zoom for high-density sampling of a suspicious range, grid for a nine-grid overview image, and transcribe for optional speech-to-text with timestamps. - Use Case: A teammate sends a screen recording of an App scrolling bug. The agent scans the video, reads the motion timeline to locate the suspicious 3-5 second window, zooms in for dense frames, and reports exactly when the panel stopped following the finger. ## Quick Start Ask the agent to look at this video and tell me what happens at each second and where the interaction goes wrong.
