The Ultimate Claude Code Research Workflow: Skills + NotebookLM + Obsidian

What happens when you stop using Claude Code, NotebookLM, and Obsidian as separate tools and instead wire them together into a single automated research pipeline? This 14-minute walkthrough from Chase (The Enterprise AI) lays out the exact architecture: Claude Code as the orchestrator, YouTube as a data source, NotebookLM as a free AI analysis engine, Obsidian as a persistent knowledge layer, and the Claude Code Skill Creator as the glue that turns the whole thing into a reusable one-command workflow.

The Architecture: Four Tools, One Pipeline

The video opens with Chase presenting the core thesis on an Excalidraw whiteboard. The pipeline flows left to right: Claude Code initiates a YouTube search via a custom skill, pipes the results into NotebookLM for analysis (across multiple Notebook LM steps), and returns the finished output back to Claude Code. Obsidian sits underneath the entire flow, capturing every markdown file the system produces.

The key insight Chase emphasizes repeatedly: this is a template, not a prescription. YouTube is his data source because he's a content creator researching video trends. But the architecture works the same way with PDFs, articles, internal documents, or any other information source. Swap the input skill, keep the pipeline structure.

Step 1: Build the YouTube Search Skill

The first skill handles data acquisition. Chase uses the Claude Code Skill Creator plugin (/plugin to install, /skill-creator to invoke) and describes what he wants in natural language: a skill that uses yt-dlp to search YouTube by query and return structured video results.

The Skill Creator automatically generates the skill file inside the .claude folder. No manual coding required. The skill wraps yt-dlp's search capabilities into something Claude Code can invoke directly.

Step 2: Connect NotebookLM via notebooklm-py

NotebookLM doesn't have a public API, so the integration relies on the open-source notebooklm-py library from GitHub (github.com/teng-lin/notebooklm-py). Installation is straightforward terminal commands followed by notebooklm login to authenticate via browser.

Once installed, the CLI gives Claude Code full programmatic access to NotebookLM's feature set:

Capability	Command
Create notebook	notebooklm create "My Research"
Add sources	notebooklm source add <url or file> (up to 50)
Chat with sources	notebooklm ask "What are the key themes?"
Generate audio	notebooklm generate audio "make it engaging"
Generate infographic	notebooklm generate infographic
Generate slide deck	notebooklm generate slide-deck
Download artifacts	notebooklm download audio ./podcast.mp3

Chase then uses the Skill Creator again to wrap this CLI into a Claude Code skill, giving the Skill Creator the entire GitHub repo as context. Claude Code understands its own skill ecosystem well enough to self-improve in this way.

Step 3: Combine Sub-Skills into a Super Skill

This is where the Skill Creator really pays off. Rather than invoking the YouTube skill and the NotebookLM skill separately, Chase creates a single "YouTube Pipeline" skill that chains them together. One /skill-creator prompt describes the full flow: search YouTube, send results to NotebookLM, run analysis, generate deliverables, return everything.

The result is a single slash command that executes the entire research workflow. No intermediate steps. No manual handoffs.

Step 4: Execute the Pipeline

Chase demonstrates the pipeline live. His prompt: search for videos about Claude Code and MCP servers, find the top five, analyze what's driving views, identify outliers and content gaps, and generate an infographic.

The pipeline runs for about six minutes. Most of the processing time is NotebookLM's analysis -- and critically, those are tokens Google pays for, not the user. The text analysis portion is fast; deliverables like infographics or slide decks take longer (up to 15 minutes for a full slide deck).

The output: a full markdown research document with key takeaways (top MCP servers: Supabase, Context7, Figma, Sentry, Shadcn), engagement analysis, content gap identification, and a generated infographic -- all saved directly into the Obsidian vault.

Step 5: The Obsidian Feedback Loop

This is the long-term play. Every research run produces markdown files inside the Obsidian vault. Those files use wiki-style double-bracket links, making them navigable both for the human (through Obsidian's graph view) and for Claude Code (which reads the markdown directly).

The CLAUDE.md file at the root of the vault acts as a "brain within the brain" -- it tells Claude Code about conventions, preferences, output formats, and work style. Over time, as more research accumulates, the user refines CLAUDE.md with a simple prompt: "update CLAUDE.md so it better reflects my work style, analysis, and output preferences based on our latest conversations."

This creates a self-improving loop. The more you run the workflow, the more data Claude Code has about how you think and what you want. The more you refine CLAUDE.md, the better the next run's output matches your expectations. Over a month, the effect is noticeable. Over a year with hundreds of documents, it's transformative.

The Bigger Pattern

Chase closes by zooming out to the generalized architecture. The specific tools (YouTube, NotebookLM) are interchangeable. What matters is the structure:

1. A workflow -- any multi-step process that helps you in your work
2. Skills -- each step wrapped as a Claude Code skill, composable into a single super-skill
3. Obsidian -- persistent markdown knowledge base that both human and AI can read
4. CLAUDE.md -- the feedback mechanism that trains Claude Code on your preferences over time

The workflow doesn't have to be research. It could be code review, competitive analysis, client briefings, or anything else that involves gathering information, processing it, and producing deliverables. The template holds.

Key Takeaways

• Claude Code's Skill Creator can chain multiple sub-skills (YouTube search, NotebookLM analysis) into a single pipeline skill that executes an entire workflow with one command
• NotebookLM integration via notebooklm-py offloads AI analysis to Google's servers, saving token costs while gaining access to deliverables like podcasts, infographics, and slide decks
• Obsidian serves as a persistent context layer -- every research output becomes part of Claude Code's growing knowledge base through transparent markdown files
• The CLAUDE.md feedback loop creates a self-improving system where Claude Code gets better at matching your preferences with every session
• The architecture is a template: swap YouTube for any data source, swap NotebookLM for any processing tool, and the pipeline pattern still works

Resources

• notebooklm-py GitHub Repository -- Open-source Python API and CLI for Google NotebookLM
• Claude Code Skill Creator -- Built-in plugin for generating Claude Code skills from natural language descriptions
• Obsidian -- Markdown-based knowledge management app with graph view and wiki-style linking
• yt-dlp -- Command-line tool for downloading and searching YouTube videos
• Chase AI Plus -- Community with free resources and skills referenced in the video