YT Pull

YouTube transcript extraction and analysis skill for Claude Code. Paste a URL, get a structured summary.

What It Does

Pulls YouTube video transcripts and metadata, then generates structured summaries via Claude — all from the terminal. Supports single videos and batch channel mode.

• Transcript extraction — structured entries with timestamps, plus clean plain text
• Metadata fetch — title, channel, date, duration, views, description
• Three analysis depths — brief, standard, deep
• Channel mode — list recent videos from a channel, batch-pull transcripts
• Auto-dependency management — creates a venv and installs packages automatically
• No video/audio download — only subtitles and metadata

Usage

Single Video

/yt-pull https://www.youtube.com/watch?v=dQw4w9WgXcQ

Supported URL formats:

• https://www.youtube.com/watch?v=VIDEO_ID
• https://youtu.be/VIDEO_ID
• Raw 11-character video ID

Channel Mode

/yt-pull --channel @nateherk --max-videos 5

Channel input formats:

• @ChannelHandle
• https://www.youtube.com/@ChannelHandle
• https://www.youtube.com/c/ChannelName

Options

Flag	Values	Default
--lang	Language code (en, es, fr, ...)	en
--output	Output directory path	./yt-pull
--depth	brief, standard, deep	standard
--max-videos	Max videos to list from channel	10

Output Structure

Single Video

yt-pull/{video-slug}/
├── transcript.txt        # Clean plain text (deduplicated, paragraphed)
├── transcript.json       # Structured [{text, start, duration}, ...]
├── metadata.json         # Title, channel, date, duration, views
└── analysis.md           # AI-generated summary & analysis

Channel Mode

yt-pull/{channel-slug}/
├── channel.json          # [{id, title, upload_date, duration, views, url}, ...]
└── {video-slug}/         # One directory per pulled video
    ├── transcript.txt
    ├── transcript.json
    ├── metadata.json
    └── analysis.md

Analysis Depths

Brief — 2-3 sentence summary + 3-5 key takeaways

Standard — Video info, 1-2 paragraph summary, 5-8 key points, notable quotes, topics covered

Deep — Everything in standard, plus: timestamped outline, argument analysis, audience context, related content, critical assessment, action items

Setup

Prerequisites

• Python 3.10+
• Dependencies are auto-installed into /tmp/yt-pull-venv/

As a Claude Code Skill

Copy SKILL.md to ~/.claude/skills/yt-pull/SKILL.md and fetch_transcript.py alongside it.

Standalone Script

# Fetch transcript + metadata
python3 fetch_transcript.py VIDEO_ID --output-dir ./output

# List channel videos
python3 fetch_transcript.py --channel @ChannelHandle --max-videos 10 --output-dir ./output

Architecture

Two-file design:

• SKILL.md — Skill definition with the full execution pipeline (parse input, fetch transcript, spawn analysis subagent, output results)
• fetch_transcript.py (~465 lines) — Standalone Python script that handles:
  • Auto-dependency installation (youtube-transcript-api, yt-dlp) via venv/uv/pipx
  • Primary transcript method: youtube-transcript-api (lightweight, structured timestamps)
  • Fallback: yt-dlp with VTT parsing and cleaning
  • Metadata via yt-dlp --dump-json
  • Channel video listing via yt-dlp --flat-playlist

The skill spawns analysis as a subagent to keep large transcripts out of the main Claude Code context window.

License

MIT