watch-locallisted

# /watch-local -- Claude watches a video using local compute You don't have a video input; this skill gives you one. A PowerShell launcher runs Python workers natively on a portable runtime the plugin provisioned for itself (yt-dlp, ffmpeg, faster-whisper -- nothing installed on the system), extracts frames, ALWAYS transcribes with faster-whisper locally, optionally compares against creator-uploaded captions, then prints frame paths + transcript. You then `Read` each frame path to see the images and combine them with the transcript to answer the user. **Compute modes (auto-detected at setup, cached in config):** with a usable NVIDIA GPU, video decode runs on NVDEC and whisper on CUDA. Without one, everything runs CPU-only (whisper on int8) -- slower but fully functional. The report's **Compute** line states which mode a run used. After driver/hardware changes, re-detect with `setup.ps1 -DetectGpu`. **Invocation note:** examples below use `powershell.exe` (Windows, ships with the OS). Always pass `-NoProfile` -- user profiles print noise and start background tasks in the captured output. On Linux/macOS hosts, replace `powershell.exe -NoProfile -ExecutionPolicy Bypass` with `pwsh -NoProfile` -- the scripts run on PowerShell 7 on any OS (CPU mode). `pwsh -NoProfile` is also a supported engine on Windows when it is installed. PowerShell 7 is NOT preinstalled on Linux/macOS: before the first invocation on those hosts, check `command -v pwsh`. If it is missing, STOP and tell the