YouTube video to skill with example frames

1. Download

• Run pip install -U yt-dlp first. An old version fails with 403.
• If you get 403 or 429, run yt-dlp --cookies-from-browser chrome -f "bv*[height=1080][ext=mp4]" -o video.mp4 URL.
• Subtitles often return 429 even with cookies. Transcribe locally instead:
  • ffmpeg -i video.mp4 -ac 1 -ar 16000 audio.wav
  • uvx --from mlx-whisper python -c "import mlx_whisper; r=mlx_whisper.transcribe('audio.wav', path_or_hf_repo='mlx-community/whisper-large-v3-turbo')"
  • Pip is blocked by PEP 668, so use uvx. The first model download takes about 8 minutes, so run it in the background.
  • Write the segments out with [m:ss] timestamps.

2. Map concepts to timestamps

• Read the transcript and list every concept with the moment its example is on screen. Leave out sponsor segments.
• Do NOT sample on a timer and do NOT ship contact sheets. The user wants the frame of the thing being explained.
• To find candidates, make small xstack grids of frames around each timestamp. drawtext may be missing from ffmpeg, so don't rely on it for labels.