I don’t really rely on a single AI tool anymore. Not necessarily because I went out of my way to build an AI-heavy setup, but because these tools have just naturally become a part of how I get work done. At this point, it’s hard to avoid AI in your workflow, so might as well make it work for you. I’ve tested a lot of AI tools by now, and while I have some on rotation, others have made their way into my productivity stack more permanently.

NotebookLM has been a staple in my study sessions for over a year now, and it’s not going anywhere soon. Setting up a local LLM is something I just wanted to give a shot because of the hype, but my gpt-oss model ended up becoming a favorite. And Claude has been more of a recent discovery for me. Each has their own strengths (and weaknesses) that handle different parts of my work and projects. And they’re even better together.

My local LLM handles exploration

It’s like a little research hub and idea playground

My model of choice is gpt-oss 20B, which I run through LM Studio. Its training data has a cutoff date of around mid-2024, which isn’t too bad. It’s recent enough that most topics I care about aren’t outdated. It’s a tad slower than my cloud models, but I like it for its matter-of-fact approach. My local AI doesn’t really learn from my behavior the same way, so I have to be more intentional with my prompts and give it a little more to work with.

I can adjust my prompts and rephrase them as much as I want since there are no paywalls or usage caps. Plus I don’t have to worry about content filtering or unwarranted censorship on perfectly reasonable queries. All of these elements create the perfect sandbox environment to throw ideas around, explore niche topics, and experiment with prompting techniques. And in the process, I take note of anything useful I’ve learned.

Local models are great for discovery and experimentation. Basically, it’s my low-pressure first layer because of how disposable the workflow is.

NotebookLM helps me make sense of my new findings

Turning the content into something usable

If you’re an avid NotebookLM user like me, you probably already know the drill: gather your sources and plug them into a notebook. Many of my sources actually come straight from my local LLM. As mentioned, I take notes during my chat session in a local notes app, and then I store the text files in a folder that’s synced to Google Drive so I can fetch them directly from within NotebookLM. Sometimes I bring the entire LLM conversation into a notebook - I use this LM Studio Conversation Converter tool to convert the JSON file into Markdown or txt. And I also bring in a lot of web links from my browser using the NotebookLM Tools extension.

Once my sources are inside the relevant notebook, I usually start with a mind map because I’m a visual learner. One of my other favorite NotebookLM features has been Studio Quizzes - they actually explain the answer regardless if you got it right or wrong. It’s also just a great tool for learning new things, including hobbies and software that have heavy documentation for rules or user guidance. Another thing I rely on quite a lot with NotebookLM is the preset prompts in the text bar - I don’t always know which questions to ask, so NotebookLM guides me in the right direction.

Claude is where things take shape

Where refinement happens

I often use Claude on its own without my local AI or NotebookLM, so it’s not always part of the AI chain. Its biggest role in my workflow is for designing interactive prototypes using the Artifacts feature. And I’m not talking about Claude Code, just the regular Claude chatbot desktop app. It can build anything you describe with natural language, including mobile screens, websites, UI components, user flows, and much more. It’s great for any graphics or design-related workflow where you need to experiment or iterate on ideas quickly, as those variations take a lot of time to create in dedicated design software.

However, Claude is also a very comprehensive chatbot. It connects to NotebookLM through MCP, so you can fetch data from a notebook directly from within Claude. It also connects to Google Drive and other Google products, so you can export your NotebookLM content and still access it from Claude even if you don’t set up the MCP server.

Claude has a big context window, so I can work with bigger sets of documents at once, and since it's not strictly tied to my sources like NotebookLM, this is the phase where I get some outside perspective. It has exceptional intent recognition, no matter how messy my prompts are. It just gets what I’m trying to convey better than most bots, which is why it’s usually my last stop in the workflow or learning process.

The difference is in how I split the work

Using these three tools together just covers my bases more than any other setup I’ve had. There’s no strict order here, and I don’t always use all three together. But the general pattern involves my local LLM for discovery and exploration, NotebookLM for structured understanding, and Claude for usable outputs. It’s not a complex system by any means or even tightly integrated, but having this AI kit in my arsenal removes a lot of the friction that comes with learning and working.