Youtube Transcript

Skill: youtube-transcript

When to invoke

User mentions:

• "get the transcript of this YouTube video"
• "summarize this YouTube link"
• "what does the speaker say at "
• pastes a youtube.com/watch?v=... or youtu.be/... URL and asks anything that requires reading what was said

How to invoke

The skill ships a uv venv at .venv/. Run:

cd <plan-root>/skills/youtube-transcript
.venv/Scripts/python.exe fetch_transcript.py VIDEO_ID

Or call from another script:

import sys
sys.path.insert(0, "<plan-root>/skills/youtube-transcript/.venv/Lib/site-packages")
from youtube_transcript_api import YouTubeTranscriptApi
api = YouTubeTranscriptApi()
fetched = api.fetch(video_id, languages=["en"])
for s in fetched:
    print(f"[{s.start:.1f}s] {s.text}")

Fallback to yt-dlp

If youtube-transcript-api fails (geofence, rate limit, etc.):

.venv/Scripts/yt-dlp \
  --write-auto-subs --skip-download \
  --sub-langs en --sub-format vtt \
  -o "%(id)s.%(ext)s" \
  "https://www.youtube.com/watch?v=VIDEO_ID"

Returns

• youtube-transcript-api: list of FetchedTranscriptSnippet objects with .text, .start, .duration
• yt-dlp: a .vtt file in current directory named after the video ID

Constraints

• Some videos disable transcripts entirely; both tools will fail. Tell user.
• Auto-generated transcripts are imperfect; flag as "auto-generated" if is_generated=True on the transcript object.
• Region-blocked videos may return 403; route through yt-dlp with --proxy if needed.

Citations

• youtube-transcript-api: https://github.com/jdepoix/youtube-transcript-api (MIT)
• yt-dlp: https://github.com/yt-dlp/yt-dlp (Unlicense)

Verified

Smoke-tested 2026-04-30 against video ID 0eKVizvYSUQ. Returned 128 segments. Skill works.