YouTube Video Analyzer

Agentic workflow for extracting and analyzing YouTube video content locally.

Prerequisites

Ensure these are installed (via Homebrew):

• yt-dlp — Video/subtitle downloader
• ffmpeg — Audio extraction
• whisper-cpp — Local transcription (+ model at ~/.local/share/whisper/ggml-base.en.bin)

Agentic Workflow

When a user provides a YouTube URL:

Step 1: Extract Transcript

Run the extraction script from this skill's directory:

${SKILL_DIR}/scripts/yt-transcript.sh "YOUTUBE_URL"

Output: Transcript saved to /tmp/yt-transcript-{video_id}.txt

Step 2: Read the Transcript

Use the Read tool to load the transcript:

Read /tmp/yt-transcript-{video_id}.txt

Step 3: Analyze Based on User Intent

User Request	Analysis Pattern
"Summarize this video"	Structured summary with key points, takeaways
"What are the main topics?"	Topic extraction with timestamps
"Tell me about X"	Search transcript for X, provide context
"Create notes"	Formatted notes with section headers
"Find quotes about Y"	Extract relevant quotations

Script Behavior

The yt-transcript.sh script automatically:

1. Fast path: Fetches existing YouTube captions (no AI needed, instant)
2. Fallback: Downloads audio + transcribes locally via whisper-cpp
3. Cleanup: Converts VTT format to clean plain text

Example Interactions

User: "Summarize this Dreamforce talk: https://youtube.com/watch?v=..." Claude: Runs script → Reads transcript → Provides structured summary

User: "What did they say about Agentforce in this video?" Claude: Runs script → Reads transcript → Searches for Agentforce mentions → Provides context