YouTube Video Analyzer
Agentic workflow for extracting and analyzing YouTube video content locally.
Prerequisites
Ensure these are installed (via Homebrew):
yt-dlp— Video/subtitle downloaderffmpeg— Audio extractionwhisper-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}.txtStep 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:
- Fast path: Fetches existing YouTube captions (no AI needed, instant)
- Fallback: Downloads audio + transcribes locally via whisper-cpp
- 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