rea-video
An Apache-2.0 agent skill that takes an existing video apart. It splits the video into shots, separates the sound into voice, music and effects, then classifies each shot: filmed, AI-generated, avatar, 3D, built in code (kinetic type, UI, charts) or screen recording. The call rests on measurements (frame cadence, sensor noise, AI-content tags in the metadata) before appearance. For each shot it returns how to remake it: shooting and lighting plan, a generation approach for an AI-made shot, or a Remotion project for code shots. The reverse-engineered prompt for each shot only comes if you plug in a model that can watch video; without one, the plan command returns drafts your agent finishes. Two outputs in the working folder: report.html for you, and VIDEO.md, an open layered format your agent reads without loading everything. It works offline and free with its own tools; an optional multimodal model, plugged in through an endpoint and a key, sharpens the call on videos recompressed by platforms. It renders no video itself.