watch-and-learn

# watch-and-learn Turn a video that *demonstrates a procedure* into a reusable Claude Code skill (or a small set of them), then prove the skill works. The video is the teacher; the output is a skill someone can invoke later without watching anything. Watching is local and free: `scripts/watch.py` extracts keyframes (PNG) + a labelled contact sheet and transcribes the audio (PyAV + faster-whisper, no API keys, no torch). See `references/watching.md` for how the engine works and its options. > **Placeholders:** `<this-skill>` = this skill's own directory (where this SKILL.md lives). `<file-or-URL>` = the video you're processing. ## Prerequisites (one-time per machine, usually already done) `watch.py` **self-locates its toolchain venv and re-execs into it** — no activation, and a plain `python3` invocation is fine (it finds the venv via `$VIDEO_WATCH_TOOLS`, a `tools/venv` up the directory tree, or `~/.video-watch-tools/venv`, which is where the bundled installer puts it). You normally do **nothing** here. Only if `watch.py` reports the toolchain is missing (a fresh machine) do you install it once: ```sh bash <this-skill>/scripts/install-tools.sh # creates the venv at ~/.video-watch-tools (av, Pillow, faster-whisper, yt-dlp) ``` First transcription downloads the Whisper `base` model (~150 MB), cached thereafter. ## Workflow ### 1. Qualify — is this worth a skill? Before spending on a full watch, decide whether the video actually teaches a **reusable, transferable pro...