What this article is about
This article compiles the full content of the free lecture "The Complete Guide to NotebookLM" (a session with over a hundred attendees). The value of NotebookLM is that it turns your data into a knowledge space you can question, compare, and organize. This article shows you, step by step, how to take it from a place to store data to a knowledge-analysis assistant that gives you action recommendations.
· People who constantly need to read piles of reports, papers, and instructional videos, can never finish them, and worry about missing the key points
· People who already drop material into NotebookLM to summarize it and want to take the next step
· Learners who want to turn multiple teachers and sources into their own AI advisor
· The four-stage evolution of reading reports, and the exact prompt for "letting the report read me"
· Real cases of a multi-source knowledge base and a cross-field collaborative syllabus
· Two hard-won data-organizing tips: the PDF trap and Markdown content boundaries
Start with positioning: the two directions of AI empowerment
AI use today splits into two very distinct directions. A study OpenAI released in September 2025 tracked the behavior of 700 million users worldwide and found that most usage today is concentrated in the first type, with only a small share in the second.
Letting AI carry out the work itself: writing articles, translating, making slide decks, generating images. This is where most people's usage sits today.
Testing ideas, supporting decisions, spotting blind spots. Fewer people work this way, but this is exactly where knowledge workers pull ahead.
NotebookLM takes the second path. Rather than waiting for AI to get more automated, the more practical route is to organize your own knowledge first, so AI is equipped to act as your advisor.
The four-stage evolution of reading reports
Back before AI, there was too much information and you often couldn't get through it all.
What most people do now. And a familiar doubt comes with it: if AI does all the summarizing, are we actually learning anything?
Have AI summarize all ten reports first to quickly decide which one to read first, then still read that one closely yourself. AI does the filtering; the close reading is still on you.
Hand the report to AI and ask it, based on my situation, to give me recommendations and an action plan that fit me.
Turning on memory mode works, but conversation memory gets mixed with errors and outdated information. A cleaner approach is to prepare a personal profile.
Basic details, core identity, positioning, career path — a quick briefing written specifically for an AI advisor.
The prompt for stage four is very simple:
Going from "I read the report" to "the report reads me" makes the very same material worth something completely different.
From one report to a whole batch: the multi-source knowledge base
If one report can give you recommendations, what about ten? What about a teacher's series of twenty or thirty videos? NotebookLM's capacity can handle this: on the early-2026 plans, a notebook holds 50 sources on the free tier and up to 300 on the paid tier (actual limits are subject to the official announcement). Once you scale up, it shifts from "using one report to read my résumé" to "using a knowledge base of 20 teachers to read two or three years of my work journals."
Cross-field collaboration is another practical use case. Back when I taught photography, I co-taught a parent-child photography course with a teacher from the body-and-mind field. I put my years of course materials in one knowledge base and he put his in another. When it was time to build the joint syllabus, I opened a new notebook, imported both sides' sources, and told the AI, "Based on these two syllabuses, help me design eight classes, an hour and a half each." Out came a joint syllabus we could actually use.
Here's a detail I learned the hard way: keep the knowledge bases maintained separately — there's no need to merge everything into one. Early on we poured both people's material into the same base, and later realized that base could only ever serve that one collaboration. If he took on other projects, my photography material would bleed in; same for me if I collaborated with someone else. Maintaining them separately, granting shared access only when needed, and importing per collaboration is the clean way to do it.
Two hard-won data-organizing tips
Tip 1: Use cloud documents as your central knowledge base, and be careful with PDFs
For source format, go with Google Docs or plain text. Most people live in the Google ecosystem, where documents connect to NotebookLM most smoothly; if you have your own knowledge-base tool (like Obsidian), the same applies — the concept is identical.
The thing to watch out for is PDFs. A PDF looks nice because, on top of the text, it hides a lot of invisible layout code, and when AI reads it, that code and all the colorful charts throw it off. Three PDFs are fine, but drop in twenty complex ones and you'll easily blow past AI's processing limit. For content that needs precise numbers, convert it to plain text or Markdown, or paste it straight into a Google Doc — all more reliable than feeding in a PDF.
Tip 2: Use Markdown to mark content boundaries
Markdown is a very simple syntax for writing. You use hash marks to distinguish top-level, second-level, and third-level headings. Its value is that it draws boundaries around content: it lets AI see clearly where one section ends, where the next begins, and which three points belong to the same level.
Take my meeting-notes prompt as an example. Once the three-layer organizing structure is laid out clearly in Markdown, AI can tell precisely where the first layer ends and where the second begins. In an age where instructions keep getting more complex, whether or not there are content boundaries makes a real difference to the results.
How you can practice: build a small source pack
Pick three documents on the same topic, create a NotebookLM notebook, and ask it in this order:
- Start with a summary: what is the core viewpoint across this material?
- Then ask for comparison: where do these sources contradict each other?
- Next, look for blind spots: my current understanding is this (write it down) — go back to the sources for supporting evidence, counterexamples, and gaps.
- Finally, attach your personal profile and ask for action recommendations: based on my situation, what does this material suggest I should do?
Once you've walked through these four steps, you'll feel the difference between a data-organizing tool and a knowledge-analysis assistant. When you publish your organized results, remember to clearly mark which parts come from the sources, which are your own organizing, and which are AI's inferences, so the answers can be traced.