NotebookLM is already one of the best tools I’ve used for research and learning. It’s source-grounded and forces you to engage with your materials instead of letting an AI carry you through the work. For anything involving documents, weblinks, or long-form content, it’s pretty hard to beat. However, pairing NotebookLM with my productivity stack has proven again and again that I can get more out of it when I’m not using it on its own.

Since I started self-hosting more of my productivity tools lately, including my local LLM in my NotebookLM workflow was bound to happen. I know it sounds contradictory, but the local LLM isn’t there to make NotebookLM private - it just happens to be a part of my stack now, and I’m using it to generate and refine materials just as I previously would have used ChatGPT or something similar. Here’s how...

It starts in LM Studio

I use my local model to obtain information

I use LM Studio for anything that involves exploring or iterating on content quickly. Because it’s local model runner, there isn’t as much lag as you’d experience with complex or long prompts in cloud-based options. It also doesn’t collect your data or retain information across sessions unless you instruct it to; the models are static. Therefore, they don’t “learn” from your behavior, making it more dependable, less prone to confirmation bias, and less likely to hallucinate or derail the conversation. Overall, this gives me faster and more reliable responses that I can learn from or expand on.

LM Studio also comes with a host of customization tools for each model. You can set the temperature, which determines how random or creative the model’s responses will be. You can adjust the response length and sampling, which shape how the model responds. Of course, it also gives you the option of adding a system prompt, so while the model won't learn from or adapt to your prompts, you can get it to focus on specific topics or assign it a persona.

Prompting works similarly to other AI chatbot tools - prompts should be clear and specific, use clear verbs and constraints, and incorporate examples if necessary. The really cool thing about LM Studio is that it lets you edit your model’s responses. This lets me remove chunks of text when I feel they’re irrelevant, add points that I feel are missing, or combine responses from multiple of my queries.

The open-source model, gpt-oss 20b by OpenAI, is my model of choice. It’s one of the best ones for general-purpose use cases, and so far it’s been great at giving me the information I need related to my design studies. But for deep research, the gpt-oss 120b would be a better option. I use it to explore a topic from multiple angles and have it generate thorough explanations for me.

Getting my local LLM conversations into NotebookLM

All you need is Google Drive

LM Studio does give you a Copy function on every prompt, so when I’m only working with a couple of responses, I honestly just copy-paste into NotebookLM (or a Google Doc, then fetch it from NotebookLM). But when I’m dealing with larger stacks of research and information, I prefer getting the whole chat into NotebookLM.

LM Studio saves chats in JSON locally. Unfortunately, you can’t change the format it saves in, but you can easily convert the files. There’s actually a tool specifically for converting your chats from LM Studio into something more readable, called LM Studio Conversation Converter. It’s browser-based, but private and client-side. This lets me convert all of my chats into something NotebookLM can read, like Markdown, plain text, or PDF.

From there, I just save the chats to a local folder that’s synced to my Google Drive, and fetch those files from within NotebookLM.

NotebookLM is the final destination

Learning from the materials I’ve gathered in LM Studio

Once I’ve got all my sources from LM Studio into NotebookLM, I interact with them just like every other set of sources in a notebook. Here, I’m not asking it to generate anything from scratch anymore, but instructing it to help me understand and work through the research gpt-oss helped me obtain.

Lately, I’ve actually started relying more on NotebookLM’s Studio features than its chat panel. Prompting it for specifics is useful because it retrieves key points in no time, but I get more use out of the Quiz, Flashcard, and Study Guide features for active recall and retention. They transform my research stack into learning materials by generating questions, which is more useful than passively skimming summaries since I’m currently doing coursework.

LM Studio for exploration, NotebookLM for understanding

LM Studio is where I explore ideas and topics quickly, and because it’s not network-dependent or cloud-dependent, it doesn’t lag, even with complex queries. Getting those materials into NotebookLM is as easy as copy-pasting them over or syncing the files to my Drive. And then it’s just a matter of making the most of NotebookLM’s capabilities to help me understand what I’m working with.