Summarise a YouTube video with Claude, GPT, or Gemini

Steps

1. Install peepshow + yt-dlp + whisper.cpp

   All three are stand-alone CLIs.

   npm install -g peepshow
   brew install yt-dlp whisper-cpp

2. Download the YouTube clip

   yt-dlp grabs the source `.mp4`. Use `-f` to pick a small format for fast extraction.

   yt-dlp -f 'best[height<=480]' -o 'clip.%(ext)s' https://www.youtube.com/watch?v=...

3. Run peepshow

   Scene-change frames + whisper.cpp transcript emerge automatically.

   peepshow ./clip.mp4 --emit json > run.json

4. Feed to your LLM

   Drag the frame folder into Claude / ChatGPT, or pipe the JSON to the API.

   peepshow ./clip.mp4 --sink obsidian   # or --sink notion / --sink slack / etc.

Why it works

YouTube videos are the canonical case for video → LLM. Native video upload to Gemini works but burns tokens proportional to clip length. peepshow's frame-extraction + whisper.cpp transcript bundle stays under a few thousand tokens for most clips — even hour-long lectures.

When it helps

• Lecture summaries and chapter generation.
• Tutorial walkthroughs where you need both visuals and dialogue.
• Compliance / review of long-form content (podcasts, conference talks).

Pitfalls

• yt-dlp is rate-limited — don't bulk-extract without spacing requests.
• Some YouTube videos are age-gated / region-locked; yt-dlp will refuse without cookies.
• Live streams need a separate flow — yt-dlp doesn't tail by default.

Works with these LLMs

Pairs with these sinks