Turn Claude Code Into a YouTube Research Agent
Harsh Puri built a research pipeline that searches YouTube, extracts transcripts, creates knowledge bases, and generates slide decks from a single terminal session. His setup combines Claude Code with NotebookLM to turn video content into structured insights in under two minutes. The entire workflow runs from command prompts, delivering research that would take hours of manual work. Read the original here.
The workflow starts with installing two key tools: yt-dlp for YouTube search and transcript extraction, and notebooklm-py for CLI access to Google NotebookLM. These handle the heavy lifting of finding videos and processing their content into searchable format.
Set Up the YouTube Research Skill
You add a custom skill to Claude Code that teaches it how to search YouTube with yt-dlp, rank results by views and relevance, and present options for selection. This skill lives in your ~/.claude/skills/yt-research/ directory and becomes part of Claude's toolkit.
Run the Research Pipeline
You give Claude a simple instruction like "Search YouTube for videos about FGF21 and push the best ones to NotebookLM. Then generate a slide deck." Claude searches YouTube, curates the top videos, creates a NotebookLM notebook, and adds all videos as sources. NotebookLM automatically extracts transcripts and makes them searchable.
The system creates a grounded knowledge base from 15+ YouTube videos that you can query across all sources simultaneously. In Puri's test, searching for FGF21 resulted in a 16-slide deck with metabolic diagrams, comparison tables, and actionable protocols, all generated from video transcript content.
What makes this approach powerful is the combination of automated curation and cross-source analysis. Instead of watching hours of videos and taking manual notes, you get structured insights derived from multiple expert sources in minutes.
“You get a grounded knowledge base from 15+ YouTube videos in under 2 minutes. Then you can generate artifacts directly from that knowledge.”
— Harsh PuriWhat Harsh Puri demonstrates here is something every marketing team can learn from. The best AI workflows are not about replacing your voice or your process. They are about finding the right framework that lets AI meet you where you are, and then building on it over time.
Ready to try this kind of thinking in your own workflow? These step-by-step recipes will walk you through related techniques you can start using today.
