| name | youtube-transcribe |
| description | Transcribe YouTube videos and extract key information using yt-dlp. Use when the user shares a YouTube URL and wants to transcribe it, summarize it, extract key points, or dump learnings from a video. Triggers on YouTube links (youtube.com/watch, youtu.be) combined with requests to "transcribe", "summarize", "watch", "extract", "key points", "notes from this video", or "what does this video say". |
YouTube Transcribe
Transcribe YouTube videos using yt-dlp subtitle extraction. Requires yt-dlp installed via Homebrew (brew install yt-dlp).
Workflow
1. Get Video Metadata
yt-dlp --print title --print channel --print description --print duration_string "<URL>"
Prints title, channel, description, and duration on separate lines.
2. Download Subtitles
Try manual subs first (higher quality), fall back to auto-generated:
yt-dlp --write-sub --sub-lang en --skip-download --sub-format vtt -o "/tmp/yt_transcript" "<URL>" 2>&1
yt-dlp --write-auto-sub --sub-lang en --skip-download --sub-format vtt -o "/tmp/yt_transcript" "<URL>"
--skip-download avoids downloading the actual video- Output:
/tmp/yt_transcript.en.vtt
- For other languages, change
--sub-lang (e.g. no for Norwegian)
List available languages if needed:
yt-dlp --list-subs "<URL>"
3. Clean VTT into Readable Text
cat /tmp/yt_transcript.en.vtt | sed '/^$/d' | sed '/^WEBVTT/d' | sed '/^Kind:/d' | sed '/^Language:/d' | sed '/^[0-9][0-9]:[0-9][0-9]:[0-9][0-9]/d' | sed '/^NOTE/d' | sed 's/<[^>]*>//g' | awk '!seen[$0]++'
This removes VTT headers, timestamps, NOTE lines, HTML tags, and deduplicates repeated lines. For long transcripts, pipe through head -N / tail -n +N to read in chunks.
4. Clean Up
rm /tmp/yt_transcript.en.vtt
Notes
yt-dlp version warnings are safe to ignore- Auto-generated subs lack punctuation and speaker labels; infer from context
- The
awk '!seen[$0]++' dedup is essential because VTT repeats lines across overlapping cue windows
