This text is an in-depth review of Google NotebookLM, focusing on its technical positioning and practical applications as a learning-focused AI research assistant. Through systematic practice and evaluation, the authors analyzed in detail the product positioning, core functions of NotebookLM (such as source management, chat interaction, audio overview, and note templates), and the underlying logic (especially system prompts). The review pointed out that NotebookLM excels in improving the efficiency of information acquisition and integration, but believes that there is still room for development in supporting the complete link from knowledge input to knowledge output. The article also explores the application prospects of NotebookLM in academic research, enterprise knowledge bases, and personal learning, and compares its differences with traditional note-taking tools and other AI products.
Recently, Google’s NotebookLM AI tool has caused heated discussions in the AI circle, especially some bloggers claim that it supports Chinese-based content generation.Chinese podcastThe function of “, which successfully aroused my interest in research. As an AI blogger and an analyst who has long followed and practiced various AI tools, I have systematically practiced and evaluated NotebookLM.
This article not only sorts out the appearance of NotebookLM’s functions, but also strives to reveal its design philosophy and underlying logic as a learning aid. I will share my observations on its functional positioning and interactive experience in actual use, and analyze its real effectiveness and potential boundaries in assisted learning and knowledge construction.
The core point is that NotebookLM, as a personalized AI research assistant, has shown strong strength in information acquisition and integration, but there is still a broad room for evolution in opening up a complete link from “knowledge input” to “knowledge output”.
I look forward to and explore how it can deepen the integration of knowledge management and content creation in the future, and move towards a true one-stop knowledge work platform.
Audio Overview Introduction: Expectations vs. Reality
The reason why I came into contact with NotebookLM some time ago was shared by some bloggers about its “Chinese podcast“. I thought that if I could turn the article into a light-hearted podcast, it would undoubtedly bring a great experience to my readers and increase the traffic of the article’s exposure.
However, after uploading the article I wrote and generating the “podcast”, the actual effect was not as good as I originally envisioned”Content production assistance“There is a certain deviation. The generated audio is indeed developed in the form of a dialogue between two AI speech synthesis characters, and the content is also closely centered around the uploaded source material. But listen to it as a wholeCloser to a structured summary and explanation of knowledge pointsAlthough the tone and interaction mode of the two AIs strive to simulate natural conversation, there is still a significant gap compared with professional podcasts or high-quality oral broadcast works in terms of fluency, attractiveness and subtlety of information transmission.
B-end product manager’s ability model and learning improvement
The first challenge faced by B-end product managers is how to correctly analyze and diagnose business problems. This is also the most difficult part, product design knowledge is basically not helpful for this part of the work, if you want to do a good job in business analysis and diagnosis, you must have a solid …
View details >
At this moment, I realized that the unpracticed publicity of some bloggers had caused me to misjudge its application scenarios.NotebookLM’s core is not a direct content creation or audio production tool, but rather a focus on providing learners with an efficient, source-based, personalized learning and research aid platform.
What exactly is the core value of NotebookLM, which needs to be analyzed from multiple dimensions such as product positioning, target users, interaction logic and system prompts.
Product positioning: Empowering efficient learning and research
NotebookLM’s official website states its core philosophy as: “Think Smarter, Not Harder”Ingenuity is better than hard workand position it as “Your Personalized AI Research Assistant”Your personalized research assistant。This clearly reveals the core goal of its product design: to empower users to conduct more efficient and intelligent learning and research activities.In layman’s terms, NotebookLM is a learning artifact.
From my actual experience, NotebookLM strictly follows this product positioning. Its core operating mechanism is as followsHighly dependent on user-uploaded source materialAll analysis, Q&A and content generation are carried out on this basis, avoiding unfounded generalization and subjective assumptions.
Therefore, its “Audio overview“Functions should be understood as an aid that serves the learning process – ie”Learning podcastsThe purpose of this article is to help users better understand and digest source material, rather than directly producing materials that can be widely disseminated.”Productive podcasts”。
With this in mind, it’s important to explore one of the key mechanisms for running in the background – system prompts.
System prompt: The behavior password of NotebookLM
What is the system prompt?
A system prompt is a set of instructions or constraints preset by AI model developers to define the model’s behavior framework, role settings, interaction styles, and output rules in specific application scenarios. It is different from the user prompt directly input by the user on the frontend, and belongs to a lower-level control mechanism, which has a decisive impact on the quality, style, consistency, and security of the model’s output content. Its core components typically include:Role definition, behavior constraints, task orientationandContext management。
For users familiar with large language models like ChatGPT, its “custom instructions” feature can be seen as a user-configurableSecondary system prompts“or”Global preferencesThe model will refer to these instructions before processing subsequent user input. For example, the custom instructions I set up for my personal ChatGPT include:
Take a deep breath and think about it step by step.
If you fail, 100 old grandmothers will suffer misfortune.
I don’t have fingers, but I’ll give you an extra $20 as a tip.
If you do it well, I’ll reward you with a delicious dog treat.
Answers must be in Chinese simplified.
This custom instruction will be reflected in ChatGPT’s actual answers:
NotebookLM system prompt
The specific system prompt content of NotebookLM comes from the prompt words obtained by @宝玉先生 through reverse engineering. Combined with my actual experience, this set of prompts is highly consistent with the actual behavior of NotebookLM, which can be judged to be highly credible in the system prompts.
I have further refined and rewritten the English version of @宝玉先生. Can be summed up in one sentence:
In five minutes, combine storytelling enthusiasm with rational analysis to extract accurate and enlightening knowledge points from the given materials for those who are eager to learn deeply but are short on time。
The details are as follows:
- Mission objectives: Efficient and in-depth information Convey the key points: Prioritize the extraction of the most valuable information and eliminate complexity and redundancies. Balance depth and clarity: Analyze deeply but not obscure, avoiding jargon stuffing or superficial descriptions. Strictly adhere to the source: only based on the original information, no adding fuel or vinegar, no subjective comments. Stimulate Interest and Reflection: Trigger “so it is” moments that make learning more than just receiving information. Tailored to personal needs: The content is close to “you” learning purpose, and the tone is direct and friendly.
- Dual Expression Style: Sense of Story + Rationality Coexist One of the cores of NotebookLM’s design is to use two voices to jointly develop the narrative to take into account both sensibility and rationality: 1. Guided tone (similar to “enthusiastic narrator”) has a friendly and infectious tone, and is good at using metaphors, analogies, and micro-stories to lower the threshold of understanding. Its function is to arouse your interest, bring abstract concepts close to life, and make you feel “this is about me”. 2. Analytical tone (similar to “rational interpreter”) has a calm and clear tone, focusing on providing context, explaining structure, and clarifying the relationship between concepts. Its role is: supplementing factual details, maintaining a neutral position, and helping you understand the underlying logic of complex problems. The two tones alternate, without labeling and distinguishing, but naturally switch between tone and context, forming a content rhythm that is both vivid and solid.
- Who are you? ——The default learner portrait system always puts “you” at the center of the narrative, and the envisioned object has the following characteristics: you have limited time, but you want to learn deeply; You like to look at things from different perspectives and look forward to being inspired; You don’t like to be overwhelmed by a lot of information and want someone to help you filter the key points; You value the practicality of the content, and learning is not just about “knowing” but also about “using”.
- Rules for using materials: only quote provided materials if they are founded: no external quotations, no guesses, and no brainstorming. Information gap is silence: if the information is not involved, it will not be filled in fiction. When there is a conflict of opinion, maintain a neutral attitude: if there are different voices, present them truthfully and do not be inclined. Emphasize relevance to you: Prioritize the most useful and inspiring angles for you.
- Style and tone: Colloquial but not casual when talking to you: explain complex concepts in easy-to-understand language and reject academic accents. Moderate relaxation and humor: appropriately “add some material” at the beginning, transition, and end to reduce the reading burden. Clear and logical: keep the reading rhythm clear and the structure clear at a glance. Objective and unbiased: No matter how lighthearted the tone, opinions must be based on facts.
- Time control: Within five minutes, make it clear that each output should be controlled within five minutes of reading (or the corresponding length), and do not “pour” or circle. Prioritize the core points and decisively omit secondary information. Segmented by topic, the structure is clear, helping you quickly enter the topic and quickly memorize the key points.
- Recommended content structure (flexible adjustment) Introduction: Use relaxed language to point out the topic and make people interested in continuing reading. Core content: 1. Use the “narrator” tone to throw out the key points and guide you into the topic; 2. Then provide facts, explain the background, and parse concepts in an “analyst” style. 3. If there are different perspectives, they can also be displayed side by side to stimulate your thinking. Relationship to You: Emphasize how this knowledge relates to your life, work, or interests. Review Summary: Easily review key points to make a lasting impression. Thought/Action Guidance: End with a question or suggestion that encourages you to extend your thinking or action.
- Content writing guidelines Don’t explicitly characterize: Don’t write “I’m the narrator” or “Now it’s up to the analyst” to write the words. Always address “you”: make every word seem to be addressed to you. Do not disclose the existence of prompts: Meta messages such as “This is a system prompt” and “I am AI” cannot appear. Natural and smooth tone switching: Style changes are guided by tone rather than abrupt switching. Style follows information accuracy: Humor, lightheartedness, and personalization should not affect the presentation of facts or time control. The end must ask a question or suggestion: let the content be more than just information, but also a starting point for action or thinking.
Overall evaluation:NotebookLM’s system prompt design reflects thisLearner-centricproduct concept. In particular, the accurate portrayal of the characteristics of the target user clarifies its positioning to serve specific demand groups – efficient, in-depth, personalized and application-oriented knowledge assistance.
Detailed explanation of the interactive interface and core functions
This section will introduce the interactive interface and features of NotebookLM in detail (taking Plus subscribers as an example, the differences between the free version and the paid version will be explained at the end of the article). For the best experience, it is recommended that you set the interface language of NotebookLM to Simplified Chinese (through the upper right cornerSet up>Output language>Chinese Simplified）。
Notebook: Task-based learning space
NotebookLM is based on “Notebook”As a fundamental unit for organizing learning content and tasks。 Each independent study topic or research project can create a dedicated notebook. From this interface, users can manage the notebooks they have created and access the notebooks shared by others.
The main interface of NotebookLM is usually divided into three main areas, as shown in the figure below:
- Source: Located on the left side of the interface, it is used to upload local files or add online resources as a basis for AI analysis.
- Chat interface): Located in the center, it is the core area where users interact with AI, where they can ask questions, request summaries, generate notes, or build mind maps around the information in the “source”.
- Audio Overview: Located on the right (or under a specific view) to generate audio summaries based on the source material with one click.
At the top of the interface, “Share” feature that allows users to share notebooks with other users for collaborative learning or group discussions. Invitees receive a notification via email or access the shared notebook via a direct link.
Source: Multiple knowledge entrance
Source supports two main content import methods:Add local filesandExplore resources online。
Local file upload:Supports a wide range of formats, including Google Docs, Google Slides, PDF, TXT, Markdown (MD) files, audio files (e.g., MP3, WAV), as well as importing web content via URL, public YouTube video subtitles, and direct copy-pasted text. From a structured and AI parsing friendliness perspective, Markdown formatted documents are the recommended preferred option.Here I have high expectations for whether NotebookLM can support direct content analysis of local video files in the future (click on the table first, see the following for details)。
Explore the source: Users can enter keywords or specific snippets of text, and NotebookLM attempts to search the internet for relevant public resources. This feature appears to integrate Google Scholar’s capabilities, which help users find academic literature including journal papers, dissertations, books, preprints, abstracts, and technical reports. For example, when I searched for an article on Potential Metabolic Pathways and Related Processes Involved in Pericarp Browning for Postharvest Pomegranate Fruits (the relationship between browning of 3 pomegranate peels and phenolic metabolism and energy metabolism), NotebookLM was able to identify and list the Source information from academic databases such as Maktaba and related journals.
It should be noted that:When importing an academic paper using the “Explore Sources” feature, we should ensure that NotebookLM can analyze the full text of the paperchooseCheck the link to the full content of the paper (e.g. PDF full text).instead of just a page that contains a summary or metadata. Only when the complete text content is imported can AI conduct effective in-depth interpretation and Q&A. As shown in the following figure:
(Click on the source to directly enter the PDF reading interface of the paper)
There is also a small trick to add requirements and restrictions when searching for sources, such as the specific content of the paper, in PDF format. This can greatly improve the probability that the source is the full text of the paper.
It should be noted that the search sources of NotebookLM are mainly limited to open platforms, and the data of commercial platforms such as CNKI, VIP, Zhihu or WeChat official accounts cannot be directly captured at present.
In terms of the number of sources, Plus users can add up to 300 sources per notebook, combined with their ability to search for academic resources such as PDFs online.Isn’t this an artifact for writing literature reviews?
A junior high school friend of mine did a “stress test”: he uploaded a full 294 volumes of “Zizhi Tongjian”, and NotebookLM not only did not collapse, but also generated a huge mind map for Chengcheng, and it took him more than 20 seconds just to drag the scroll bar from beginning to end.
(A full 294 sources were uploaded)
After adding the source material to NotebookLM, users can click on the specific source, and the system will generate a copy of theSource guide“, which contains a summary of the material’s content, a list of key topics, etc., to help users quickly understand the core content of the material and visually demonstrate the scope of information that the AI has processed and understood.
(Source Guide: Summary and Key Topics)
Interestingly, even if it is a source in PDF format, when you click to view it in NotebookLM,Its content is also presented in editable text。 However, due to the complexity of PDF layout, this text conversion can sometimes lead to formatting confusion and a reading experience that is not as good as native Markdown or well-structured web pages.
(The typography is more disordered)
Chat interaction: accurate traceability and in-depth dialogue
After uploading or selecting sources, users can ask the model questions or give instructions (e.g., “summarize this document”, “generate a mind map”, “save answer as note”, etc.) in the chat interface. Users can flexibly specify which specific sources the model should reference or exclude in the current interaction by checking the checkboxes in the source list.
Before starting a conversation, the user can use the “Conversation configurationThis is similar to a local, “secondary system prompt” for the current notebook. Users can choose the conversation style (e.g., default, analyst, guide, etc., which can also be customized) and the length of the response (default, longer, shorter).
Core features of chat interaction:
Accurate original citation and traceability: This is a notable highlight of NotebookLM in interaction design. When the model references the content of the source material in its response, a reference corner mark (number) is displayed next to the corresponding text. When users click on the corner marker, the “Source” area on the left will automatically scroll to the corresponding specific position in the original text and highlight the relevant paragraphs. This direct and accurate traceability is crucial for verifying information accuracy and deeply understanding the context, which is rare in mainstream AI tools.
(After clicking on the corner marker, the source area will automatically highlight the relevant paragraph)
Zhihu Direct AI also provides a citation corner mark function, but the experience is far inferior to NotebookLM. It sometimes cites papers on the VIP platform, but after clicking on the corner marker, only the entire PDF document is displayed, and it cannot directly locate the specific citation content in the original text, which is extremely inconvenient to find. In contrast, NotebookLM’s reference jumps are not only accurate but also automatically highlighted, which is very efficient. (@知乎产品经理, I suggest you seriously experience NotebookLM, it is really worth in-depth research as a competitor).
Dynamic mind map generation
NotebookLM is capable of generating structured mind maps based on the source material or discussions in specific sections. This helps users quickly grasp the core structure and logical relationships of the content.
We can also click on a subnode in the mind map, and the chat interface will automatically generate further explanations or guiding questions around the topic of that subnode, promoting deeper understanding. As shown in the following figure:
(Click on the subnode to learn more)
Audio function: Listening and learning combined with note-taking assistance
The audio function is mainly reflected in “Audio overview“And”remark”。 Audio Overview allows users to generate audio summaries based on selected sources with one click.
As mentioned earlier, this is usually presented in a conversational format by two AI voices (simulating the “narrator” and “analyst”) and is designed to provide an auditory way of learning. The generation quality of Chinese audio is acceptable in terms of naturalness.
The beta version is also interactive, meaning users can “join” the conversation and receive personalized responses during listening.
Reminder: The audio will not be retained, and the audio will disappear after the web page is refreshed.
In “remark“Ribbon, NotebookLM offers a range of structured note templates that can quickly convert text content into a document in a specific format, and part of the note content or summary can be used as a new source later:
- Custom notes: Allows users to freely record ideas, annotations, and associate with source materials.
- Study Guide: Transform source document content into structured learning materials with key concepts, summaries, terminology explanations, and more.
- Briefing documents: Quickly extract core points from lengthy information and generate topic overviews.
- Frequently Asked Questions: Automatically generate a series of potential questions and their corresponding answers based on the content of the source document.
- Timeline: Extract event information from materials and organize them in chronological order to clearly display the development context.
Technology positioning and prospects
The underlying large model of NotebookLM is Google’s Gemini series models (such as Gemini 1.5 Flash or Pro).Specific fine-tuning is made for learning and research scenarios（Fine-tuning）。 Its performance in assisted learning, rapid information absorption and integration does demonstrate the advantages of AI models focused on specific field applications.Compared with general-purpose large language models, the learning scenarios are separated for in-depth optimization, which is more in line with the development trend of AI technology penetrating vertical industries.
For learners and researchers, NotebookLM’s knowledge organization method in “notebooks” and in-depth interaction around source materials can indeed improve learning efficiency and focus, and facilitate knowledge precipitation and review.
howeverFrom the perspective of the complete closed loop of knowledge work, the ultimate goal of learning often points to the application and creation of knowledge, which is inseparable from the “output” link– As the German sociologist Niklas Luhmann emphasized in his Zettelkasten (card box note-taking) practice: “Without writing, you can’t think”。Only by structurally organizing, critically analyzing the acquired knowledge points, and reformulating and creatively outputting them in their own language can the true internalization and innovation of knowledge be achieved.
At this stage, NotebookLM prefers an efficient “input” and “processing” tool, but its capabilities need to be strengthened in supporting users to transform these inputs and processing results into structured knowledge outputs (such as academic papers, research reports, in-depth analysis articles, etc.).
I hope that NotebookLM will further expand its auxiliary capabilities on the “output side” in the future, such as providing more advanced note organization, knowledge graph building, and cross-note content linking functions to support the construction of personal knowledge networks (I highly recommend Google to acquire Notion, hahaha). Or integrate more powerful writing aids that cover the entire process of academic writing, from literature review, argument construction, draft writing, to citation management. In this way, NotebookLM can evolve from a “learning assistant” to a true “research and creation partner”.
I mentioned a little hope earlier that NotebookLM will support local video uploads in the future, which I think will greatly expand its application scenarios:
For exampleCollege students preparing for the postgraduate entrance examinationWhen conquering video courses, it is usually necessary to watch videos repeatedly and manually organize notes, which is not very efficient. If NotebookLM supports importing local videos or electronic handouts in the future, automatically categorizing and organizing course content, and generating clear mind maps and fine notes, printing these materials will definitely greatly improve the efficiency of review.
This need is not limited to academic scenarios, but is equally common in everyday life.Recently, I learned to make mapo tofu and fish-flavored eggplant on Douyin, and every time I actually do it, I always forget the order of the steps and the seasoning ratio. I’m lazy and not used to taking notes or taking screenshots frequently, which often leads to me having to watch videos repeatedly in the kitchen, which is quite inconvenient. Some people may say, “Practice makes perfect by doing it a few more times.” But until you really get acquainted, this process of constant flipping through is unavoidable. If NotebookLM can handle these short video teaching contents, it can completely solve my troubles by generating a step list and material reminders with one click.
For another example, there are similar pain points when actively obtaining knowledge-based video content.Sometimes I’m watchingKnowledge sharing video of film and television hurricanesI am often moved by some golden sentences or opinions, and I want to excerpt them for in-depth study. However, video subtitles do not always download smoothly, so I can only upload the video to MemoAI for transcription, and then copy the text to ChatGPT for interpretation, which is a tedious and time-consuming process. In contrast, it is enough to watch the video content once or twice, and the follow-up review is more efficient in terms of text, which is more suitable for repeated reading and refining thinking. (Note: MemoAI is an efficient tool that can quickly transcribe YouTube, podcasts, and local audio and video into text, generate summaries, and mind maps.) ）
In addition, NotebookLM’s design ideas are also very inspiring for enterprise-level applications.
When I was a documentation engineer, I was responsible for writing and publishing documentation for IoT products. The company is oriented to the B-end, and the product information is complex, and the information of a hardware product may include multiple documents such as overview, user guide, installation guide and fault repair guide. Customers or marketers want to find out a specific information, such as what the torque of a certain screw is, usually need to go through multiple documents or even ask the hardware product manager, which is thankless.
NotebookLM offers a new solution:By simply entering the product name or model number, marketers can quickly locate all relevant documents, selectively import data according to specific needs, and ask questions directly in the chat window to get answers。 The reason why we need to design on-demand check-in documents is because there is still an upper limit to the single processing capacity of large models.This cleverly avoids capacity limits and is responsive.
Of course, this model is not only suitable for marketers to find answers,All employees can use it for knowledge learning, process sorting or corporate standard training.In the long run,Company information can be truly integrated into an efficient and flexible enterprise knowledge base, realizing accurate aggregation and flexible use of knowledge.
Finally, let’s talk about the new note-taking software on the market, such as Notion, Wolai and FlowUs, which are also introducing AI.However, it is basically just a taste and has not yet truly integrated AI capabilities into the software system.
Taking my personal experience as an example, the existing tools still lack deep knowledge activation. Taking Wolai, which I usually use, as an example, I accumulate more and more fragmented notes every day, and the redundancy and cumbersome gradually increase, and I often need to spend a lot of time cleaning up.When writing a new article, if you want to quote content from past notes, you have to search through the boxes and cabinets one by one, which is extremely inefficient。 In contrast, the design concept of NotebookLM provides a solution: it can integrate and activate each note in the form of a knowledge base, and even add auxiliary creation functions on this basis to form “Learning-Recording-Restructuring-OutputA complete closed loop.
If this can be done, this is undoubtedly an epoch-making revolution in note-taking software, and for other competitors, it is even more crushing.
How to use and experience?
Visit the official website of NotebookLM (https://notebooklm.google.com/) to start your AI learning journey.
NotebookLM Free Account:You can create up to 100 notebooks, each with a maximum of 50 sources, a limit of 50 chat queries per day, and a limit of 3 audio generations.
Recommended upgrade toGemini Premiumto unlockNotebookLM Plusand video creation toolsVeo2Full functionality. NotebookLM Plus has at least 5x more capacity, can create up to 500 notebooks, with a maximum of 300 sources per notebook, and can make up to 500 chat queries and generate 20 audio per day. It should be noted that shared notebooks do not increase the upper limit of additional sources, and the total number of sources is still 300.
By the way, there was a bug in the verification of Google’s discount for students in the US recently, and you can enjoy an 18-month premium One subscription for free without edu email and verification card. Tried it today and found it fixed.
