Static Programming

Reel Coach turns any reel of a physical activity into a personal coach. Millions of people learn workouts, dance steps, yoga and sport moves from short videos, but all they can do is watch, copy and guess whether they got it right. Form-check apps cover only a fixed list of exercises someone programmed in advance. In Reel Coach, any reel is the lesson, with no setup for each activity. You upload a reel. The app reads the instructor's skeleton, splits the video into moves, picks one clean rep and pulls out the instructor's own tips, including on-screen captions. You then step in front of your camera. Your joints turn green as you match the reel's starting pose, and after a 3-2-1 countdown the instructor's skeleton is drawn on your body, one step ahead, showing where to move next. In Follow mode the guide waits for you and never runs on a clock. Reps are counted live, and spoken cues arrive just before the moment you went wrong last time, with an encouragement when you get it right. When you stop, you get 2–3 corrections straight away, each shown on side-by-side frames of the reel and you with the off joint highlighted and explained aloud in the instructor's terms. A scrubbable player lines the two videos up moment by moment. The coach remembers earlier attempts and tells you what improved. It works across activities. For a dance reel with several people, you tap the person you want to learn, and the app re-reads only that dancer. A fixed-speed mode (0.5×, 0.75×, 1×) makes you keep the reel's tempo and reports whether you were ahead of or behind the beat, a ladder up to full speed. The core design principle is that code measures and AI explains. Both videos become joint angles, not pixels, so height and distance from the camera don't matter. Dynamic time warping matches moments, not seconds, so moving slower isn't counted as a mistake. Each joint is weighted by how much it moves in the reel, so the reel itself decides what matters. The language model never invents a correction: every tip must trace back to a measured difference, and code enforces that. Pose tracking runs in the browser at about 15 fps, and live cues are written in advance, so there is no AI wait in the middle of a rep.