Turn YouTube videos into cited research notes
20 videos, cited answers, audio overviews, and a knowledge graph in your vault. Task: 20 videos, cited answers, audio overviews, and a knowledge graph in your vault Your research is trapped inside NotebookLM. Mine comes back as a knowledge graph. NotebookLM reads 300 sources. But then what? You copy-paste answers into your notes and lose the citations. The research stays in a browser window you’ll never reopen. So I built a Claude Code skill that pulls everything out. Three terminal commands - search YouTube, add 20 videos as sources, ask questions. Every answer comes back with citations pointing to exactly which video said what. Then it all lands in Obsidian as a knowledge graph. Each source becomes a file. Each topic becomes a file. Citations create links between them. I open “Claude Code” as a topic and see which 6 videos mentioned it, what they said, and follow the links to I tried connecting NotebookLM to my research months ago and gave up. Too slow. Adding sources manually in the browser, one by one, clicking through the UI. NotebookLM supports up to 300 sources - that’s the number that makes it different from just asking Claude - but who’s going to drag 300 files into a browser window? Thanks for reading! Subscribe for free to receive new posts and support my work. Now it’s different. I built a Claude Code skill that does everything from terminal. 20 YouTube videos become a knowledge graph I can walk through. Watch the full walkthrough (33 min) Before: 20 browser tabs, copy-paste, lost sources. After: one terminal command. NotebookLM is Google’s research tool. You give it sources - YouTube videos, PDFs, web articles, your own notes - and it reads all of them. When you ask a question, it gives you a cited answer pointing to exactly which source said what. No hallucination, because it only uses what you gave it. The key difference from the browser workflow: I have much more granular control over the sources. Instead of NotebookLM’s built-in suggestions that recommend whatever it thinks is relevant, I pick exactly what goes in. I can search YouTube from terminal, see the results, and choose which videos to add. This way you increase the signal-to-noise ratio. Searching YouTube from terminal with ytcli - I pick the exact videos that become research sources You can set a custom persona per notebook - tell NotebookLM to respond concise and direct, or in whatever style fits your research. The Research Workflow I wanted to research what other people are building with Claude Code and Obsidian. Instead of watching 20 YouTube videos at 2x speed and forgetting what video 2 said by video 6, I ran three commands: Search YouTube for relevant videos Search YouTube for relevant videos Create a NotebookLM notebook and add all 20 as sources Create a NotebookLM notebook and add all 20 as sources Ask questions and get cited answers Ask questions and get cited answers One terminal. No browser tabs. Claude Code orchestrated the whole thing. Watch: Live demo from terminal The citations are what make this useful. When NotebookLM says “three creators mentioned using Obsidian Bases for context management,” it points you to exactly which video, which timestamp. I tested the citation accuracy: about 60% were strong matches to exact passages, 31% partial matches, 10-15% weak. Not perfect, but way better than my memory after watching 20 videos. Watch: Citation accuracy audit Under the hood, every answer comes back as structured JSON - each claim has a marker like [1] or [2] that traces to a specific source and passage: How citations work: each claim traces back to the exact source and passage In practice, this means you can click through to the original video and timestamp: Citation markers in practice - answers point to exact sources with passage-level deep links Every answer is grounded - you can trace claims back to the source. Good enough to trust as a research starting point, not good enough to cite in a paper without checking. Here’s what a cited answer looks like when it lands in Obsidian - the exact passage from the source, with a link back to the ground truth: A cited answer in Obsidian - exact passage from the source, linked back to the ground truth The Knowledge Graph Everything comes back to Obsidian. Not as a flat dump, but as a linked knowledge graph.; tools: Claude Code. Human role: The original post does not say how a person stayed involved.; limitations: Prerequisites include Claude Code, an Obsidian vault, and a Google account. The setup takes about 15 minutes. Citation accuracy is imperfect, requiring human review for critical claims. The workflow relies on terminal commands and specific skills, which may require technical proficiency. Source reliability is secondary, as the post is a personal blog entry..
Tools
Published on Agent Work Atlas:
Overview
20 videos, cited answers, audio overviews, and a knowledge graph in your vault
Workflow
20 videos, cited answers, audio overviews, and a knowledge graph in your vault Your research is trapped inside NotebookLM. Mine comes back as a knowledge graph. NotebookLM reads 300 sources. But then what? You copy-paste answers into your notes and lose the citations. The research stays in a browser window you’ll never reopen. So I built a Claude Code skill that pulls everything out. Three terminal commands - search YouTube, add 20 videos as sources, ask questions. Every answer comes back with citations pointing to exactly which video said what. Then it all lands in Obsidian as a knowledge graph. Each source becomes a file. Each topic becomes a file. Citations create links between them. I open “Claude Code” as a topic and see which 6 videos mentioned it, what they said, and follow the links to I tried connecting NotebookLM to my research months ago and gave up. Too slow. Adding sources manually in the browser, one by one, clicking through the UI. NotebookLM supports up to 300 sources - that’s the number that makes it different from just asking Claude - but who’s going to drag 300 files into a browser window? Thanks for reading! Subscribe for free to receive new posts and support my work. Now it’s different. I built a Claude Code skill that does everything from terminal. 20 YouTube videos become a knowledge graph I can walk through. Watch the full walkthrough (33 min) Before: 20 browser tabs, copy-paste, lost sources. After: one terminal command. NotebookLM is Google’s research tool. You give it sources - YouTube videos, PDFs, web articles, your own notes - and it reads all of them. When you ask a question, it gives you a cited answer pointing to exactly which source said what. No hallucination, because it only uses what you gave it. The key difference from the browser workflow: I have much more granular control over the sources. Instead of NotebookLM’s built-in suggestions that recommend whatever it thinks is relevant, I pick exactly what goes in. I can search YouTube from terminal, see the results, and choose which videos to add. This way you increase the signal-to-noise ratio. Searching YouTube from terminal with ytcli - I pick the exact videos that become research sources You can set a custom persona per notebook - tell NotebookLM to respond concise and direct, or in whatever style fits your research. The Research Workflow I wanted to research what other people are building with Claude Code and Obsidian. Instead of watching 20 YouTube videos at 2x speed and forgetting what video 2 said by video 6, I ran three commands: Search YouTube for relevant videos Search YouTube for relevant videos Create a NotebookLM notebook and add all 20 as sources Create a NotebookLM notebook and add all 20 as sources Ask questions and get cited answers Ask questions and get cited answers One terminal. No browser tabs. Claude Code orchestrated the whole thing. Watch: Live demo from terminal The citations are what make this useful. When NotebookLM says “three creators mentioned using Obsidian Bases for context management,” it points you to exactly which video, which timestamp. I tested the citation accuracy: about 60% were strong matches to exact passages, 31% partial matches, 10-15% weak. Not perfect, but way better than my memory after watching 20 videos. Watch: Citation accuracy audit Under the hood, every answer comes back as structured JSON - each claim has a marker like [1] or [2] that traces to a specific source and passage: How citations work: each claim traces back to the exact source and passage In practice, this means you can click through to the original video and timestamp: Citation markers in practice - answers point to exact sources with passage-level deep links Every answer is grounded - you can trace claims back to the source. Good enough to trust as a research starting point, not good enough to cite in a paper without checking. Here’s what a cited answer looks like when it lands in Obsidian - the exact passage from the source, with a link back to the ground truth: A cited answer in Obsidian - exact passage from the source, linked back to the ground truth The Knowledge Graph Everything comes back to Obsidian. Not as a flat dump, but as a linked knowledge graph.
Reported result
The workflow successfully converts 20 YouTube videos into a navigable knowledge graph in Obsidian. Citation accuracy testing showed approximately 60% strong matches to exact passages, 31% partial matches, and 10-15% weak matches. The author notes this is sufficient for research starting points but requires manual verification for academic or formal citations.
Human role
The original post does not say how a person stayed involved.
Limitations
Prerequisites include Claude Code, an Obsidian vault, and a Google account. The setup takes about 15 minutes. Citation accuracy is imperfect, requiring human review for critical claims. The workflow relies on terminal commands and specific skills, which may require technical proficiency. Source reliability is secondary, as the post is a personal blog entry.
Try it
Act as a research assistant using Claude Code, NotebookLM, and Obsidian. My goal is to convert [replaceable: list of YouTube video URLs or search topic] into a cited knowledge graph in my Obsidian vault. First, perform a dry run with [replaceable: 2 sample videos] to verify the pipeline. Use the 'notebooklm' skill to add sources and generate a JSON response for the query '[replaceable: specific research question]'. Ensure citations link to exact timestamps. If the dry run succeeds, process the full list. Output the results as Markdown files in [replaceable: Obsidian folder path], creating wikilinks between sources and topics. Do not publish or sync until I confirm the citation accuracy of the sample. If inputs are missing, ask for the video list and target folder.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.