Most people treat research as a manual task.
You open 10 tabs. You watch videos. You read articles. You take notes somewhere. An hour later you have a pile of information you're not sure what to do with.
There is a better way.
This is a step-by-step guide to building a research workflow using Claude Code, NotebookLM, and Obsidian that can investigate any topic - market dynamics, emerging technologies, crypto ecosystems, content niches, anything - and gets sharper every time you use it.
Setup time: under 30 minutes
The Stack and Why It Works
Four tools. Each one handles a different layer of the problem.
- Claude Code - the execution engine. It runs commands, calls skills, manages files, and orchestrates the entire pipeline. You talk to it in plain language, it does the work.
- Skill Creator - the customization layer. A Claude Code plugin that lets you build reusable skills in natural language. You describe what you want, it generates the code and installs the skill. No programming required.
- NotebookLM - the analysis engine. Google's AI research tool that reads your sources and generates deep analysis, summaries, infographics, flashcards, podcast scripts, and more. When Claude Code offloads processing to NotebookLM, it's using Google's compute, not your Claude tokens.
- Obsidian - the memory layer. A local markdown-based knowledge system that stores everything the workflow produces. Over time Claude Code reads these files and learns how you think, what you care about, and how you want your analysis delivered.
Combined: a research system that executes on command, analyzes at scale, and improves with use.
Step 1: Install the Skill Creator
Open Claude Code. Make sure you are inside your Obsidian vault folder - this is important for Obsidian to pick up the files Claude Code generates.
Run this command:
Search for skill-creator. Install it. Exit Claude Code. Restart Claude Code.
You now have the ability to create any skill by describing it in plain language
Step 2: Create the YouTube Search Skill
This skill is what allows Claude Code to search YouTube and pull structured video data - titles, channels, subscriber counts, view counts, upload dates, URLs, and engagement ratios.
Run this command inside Claude Code:
Claude Code will generate the skill, install it, and confirm. You now have `/yt-search` available as a command.
Note: yt-dlp needs to be installed on your machine. If you don't have it
Step 3: Install NotebookLM-py
NotebookLM does not have a public API. To connect Claude Code to NotebookLM, we use an open-source project called notebooklm-py.
Repository: github. com/teng-lin/notebooklm-py
Run these commands in your terminal (not inside Claude Code - open a separate terminal window):
Then authenticate:
A browser window will open. Log into your Google account. Done. The connection is established
Step 4: Create the NotebookLM Skill
Now you need to teach Claude Code how to use notebooklm-py. Run this inside Claude Code:
This gives Claude Code a full NotebookLM skill with commands for every action NotebookLM supports - up to 50 sources per notebook, all deliverable types.
Step 5: Combine Everything Into One Pipeline Skill
This is where the workflow becomes genuinely powerful.
Instead of manually running the YouTube search, then sending results to NotebookLM, then requesting analysis - you build one skill that does all of it in sequence on a single command.
Run this inside Claude Code:
Running the Workflow
With the pipeline skill installed, this is what an actual research session looks like.
The topic: AI agent frameworks. What's actually gaining traction in 2026, what's overhyped, and where the gaps are in existing coverage.
Claude Code starts the pipeline. It calls the YouTube search skill, finds 10 videos across framework tutorials, comparisons, and developer takes - passes the URLs to NotebookLM, creates a notebook, runs analysis, and requests an infographic.
Total processing time: around 6 minutes.
Most of that time is NotebookLM processing on Google's servers - not your Claude tokens.
The result comes back as:
1. A full analysis covering which frameworks are rising vs. plateauing, what developers are actually complaining about, engagement outliers, and content gaps nobody has covered yet
1. An infographic mapping the AI agent framework landscape
1. A markdown file saved directly into your Obsidian vault with everything structured and linked - ready to reference in future research sessions
Where Obsidian Makes It a Different Tool Entirely
Everything above works as a one-time research task.
Obsidian is what turns it into something that compounds.
Every markdown file the workflow produces lands in your Obsidian vault. Over time your vault becomes a structured corpus of everything you've researched - topics, sources, analysis, patterns, conclusions.
Claude Code can read all of these files. It sees how they're linked. It understands what topics you return to, what analysis you found useful, what format you prefer.
The `claude.md` file inside your vault is where this becomes explicit. It's a configuration file that tells Claude Code how to work with you - your conventions, your output preferences, how you want things structured.
You update it by saying:
Claude Code reads the recent session, identifies your patterns, and updates the file.
Do this once a week. After a month the workflow knows you well enough that outputs start matching what you actually want without extensive prompting.
After a year - if you're doing this consistently - you have a research system that has absorbed hundreds of sessions, understands your thinking style, and operates as a trained assistant rather than a blank tool.
The Modular Point Nobody Mentions
The YouTube source is not the point.
The pipeline structure is the point.
You can replace YouTube with any data source Claude Code can access:
- PDFs - academic papers, industry reports, whitepapers
- Public web pages - news articles, documentation, blog posts
- Local files - your own notes, exported data, transcripts
- Google Drive - documents and spreadsheets you already have
The workflow template stays the same. Swap the source, keep the structure.
Research a crypto ecosystem using whitepapers and public documentation. Analyze an emerging technology using conference talks on YouTube. Map a content niche by analyzing what's performing. Study market dynamics using public reports.
Whatever the use case - the pipeline, the analysis layer, and the memory system remain identical.
What You End Up With
A research system that:
- Executes full research pipelines on a single command
- Offloads heavy analysis to Google's infrastructure via NotebookLM
- Produces structured deliverables - infographics, mindmaps, audio, flashcards - automatically
- Saves every result to a local knowledge base
- Learns your preferences over time and improves its outputs accordingly
The 30-minute setup pays for itself the first time you use it
