wtf is watch-skill?

Watch Skill is a tool that gives AI coding agents the ability to watch video, understand what happened in it, and check their own work by recording and reviewing it. It turns videos, live streams, meeting recordings, and screen recordings of an agent's own actions into a searchable, timestamped record that an agent can search, ask questions about, and cite specific moments from. The project describes three main abilities it adds to an agent. Watch lets it process video from more than 1,800 websites, live streams, local video files, and recordings of browser or desktop sessions, extracting frames, on screen text, and transcripts. Remember builds a persistent, searchable index of everything analyzed, so an agent can ask a new question about a video later without re-downloading or re-processing it, and can find related material across every video it has seen before. Verify, referred to in the README as THE LOOP, is meant for when the video is a recording of the agent's own work: it records the session, checks the result against plain language criteria the user sets, guides a fix if something is wrong, and produces before and after evidence that the fix worked. It is installed either as a plugin inside Claude Code, or through an install script on macOS, Linux, or Windows, and is also usable through a command line tool, a REST API, and 23 tools exposed over the MCP protocol so many different AI coding agents and editors can use it. Transcription and search can run entirely on a local machine without any API key, while answering visual questions about a video requires connecting a vision capable AI provider such as Anthropic, Gemini, OpenAI, OpenRouter, or a locally run Ollama model. It is written mostly in Python, released under the MIT license, and the project includes its own published cost and accuracy benchmarks rather than relying on unverified marketing claims.