'Watch This Video for Me' Consolidates Into a Standard Agent Skill, With at Least Five Independent Implementations Now Public
A widely shared August 3 post announced a /watch-video skill covering YouTube, Loom, Zoom recordings, Vimeo and local files that pulls the transcript, extracts frames, and runs a vision pass over key moments. Verification turned up at least five independently built public repos doing the same thing (alexlarcheveque/claude-watch, Newuxtreme/watch-video-skill, jordanrendric/claude-video-vision, bradautomates/claude-video, taoufik123-collab/claude-watch), all converging on the same pipeline of yt-dlp for retrieval, ffmpeg for frame slicing, and a timestamped frame-plus-transcript pair handed to the model. The convergence is the signal: when five builders independently land on the same architecture in the same quarter, the pattern is stable enough to copy rather than design.
Source
↳ Follow the thread