Notion Agent Complete Practical Tutorial - Creating Intelligent Workflows with AI

In 2022, when the world was just beginning to feel the reality of AI, Notion was already one of the first tools to integrate GPT-3. Today, it has been building its AI feature layout for many years. Therefore, if you ask which "note-taking software" has the deepest integration with AI and the greatest practical value, I would recommend Notion without hesitation.

However, over these 3 years, Notion AI has evolved through countless iterations, but I rarely see people discussing the usability and power of Notion AI in my timeline, which makes me feel somewhat regretful. Model iterations can easily stir up waves of enthusiasm, but after calming down, whether AI can truly optimize or even rewrite outdated workflows, rather than becoming a password for major accounts to devour traffic, is what I truly care about.

Therefore, in today's article, I want to share the following topics:

1. Why Notion AI is worth trying
2. How much potential Notion Agent has
3. How I use Notion AI
4. Subscription advice for Notion AI

Every time I write about Notion, I can't quite control the length. The full article is very long, but I'm convinced these are details that few people would write about. Next, I'll briefly introduce the basic capabilities of Notion AI first. If you want to see the core content of this article about Notion Agent, you can start reading from the second section.

1. Basic Capabilities of Notion AI

First, like all the AI tools you've used, the basic way to interact with Notion AI is also through Q&A-style chat. You can select a paragraph on the page and ask questions directly, or you can have longer discussions in the AI panel that expands on the right side, as shown below.

Additionally, the content of the selected paragraph is automatically added to the context information in the right panel, so you don't need the extra step of copying and pasting.

Besides Q&A for specific content, Notion AI can also perform semantic searches across your entire workspace. When you only remember the approximate content of a note but forget what the title is or which database it exists in, you can directly ask AI with vague descriptions, and AI can help you find the location of relevant notes.

Based on this feature, I've built an item management center in my Notion system. Some infrequently used important items have storage locations set, so I can ask like this:

Moreover, because Notion AI has integrated the latest models from Anthropic, OpenAI, and Google, it possesses multimodal processing capabilities. It can process text information, analyze uploaded CSV, PDF, and image files, search the web, and you can also manually send it webpage links, then have it read the webpage text and answer.

After generating an answer, Notion AI can directly perform CRUD operations on your pages or databases. Therefore, you can have Notion AI directly modify the original note text, or you can require AI to create a new page and then save the generated answer in a specified location (including databases).

Basically, questions that DeepSeek or Doubao can answer, Notion can answer too, and usually can answer better. But the real key here is that at all times when you're writing, taking notes, summarizing, reflecting, and reviewing, when you need AI, you don't need to open a second AI tool. All AI needs don't require leaving Notion, truly achieving All in One.

In note-taking scenarios without integrated AI, you usually need to switch back and forth between multiple windows. If you need multiple rounds of dialogue or need to reference multiple note contents, this number of jumps increases exponentially. And when you want to save AI-generated answers, you need to reorganize the format, add tags, and delete excess content again. These operations all need to be completed manually.

This cumbersome process not only wastes time but, more importantly, interrupts your flow state. Because while waiting for answers to generate, you often get distracted doing other things or scrolling through your phone watching videos, and then ten minutes quietly slip away.

The convenience of Notion AI lies not only in its ability to answer your questions but also in its ability to process all information "inside and outside" your workspace in a one-stop manner. You no longer need to copy a note paragraph for Doubao to read, nor do you need to transport ChatGPT-generated answers back to your note-taking software. One less jump makes a world of difference in product experience.

When you need to write daily or weekly reports, you can directly reference all the documents you've written this week in the Notion AI window, then have it directly read and generate a report. When you need to review the key decision-making process of a project, you can reference multiple meeting records to have AI extract all discussed solutions, or directly read the page's edit history and analyze the reasons behind decisions.

You can also use the Save to Notion plugin to build a web clipper database in Notion, then create an AI field. In this field, you can preset specific information processing Prompts, such as generating reading summaries or extracting specific key information. This way, without subscribing to Readwise, you can also have an unlimited AI-assisted reading center.

Notion is inherently suitable for organizing high information density and high-value content. Combined with the latest AI models and the fastest AI interaction methods, it naturally achieves better results than other note-taking software.

Moreover, thanks to Notion's 10 years of open ecosystem development, you can also directly search connected external information sources like Google Drive, Google Calendar, GitHub, Gmail, etc. in the AI window. Of course, the prerequisite is that you need to be within these ecosystems.

Besides these built-in "connectors," you can also establish connections with more tools through MCP, such as Cursor, Manus, Perplexity, ChatGPT, etc. You can send Notion page links to these tools, and they can directly read note content without cumbersome copy-paste operations. These tools can also directly modify Notion pages based on your needs.

For example, when you've obtained a lengthy research report through Manus, in the past you might have needed to manually copy and paste it into other note-taking software. But now you can send the Notion page link to Manus, and Manus can directly help you write it to a specified location in Notion.

Notion AI can fully take over your information processing workflow. Your notes, your tasks, your project documents, and even your connected third-party tool data can all be processed more quickly and directly. But these are just basic applications of Notion AI. If it's only rapid invocation of "chat window + information sources," then Notion AI isn't enough for me to use long-term.

The true purpose of this article is to reveal to you the key capabilities and potential of Notion Agent.

2. Notion Agent

What is an Agent

Simply put, an Agent is an AI system that can autonomously complete multi-step tasks on behalf of humans.

Most AI Q&A tools you've used can only provide information or ideas. After getting answers, you still need to manually execute them. Moreover, they don't know who you are, don't know what projects you're working on, don't know your work habits and preferences. Every conversation starts from zero, and you need to repeatedly explain background and describe requirements. Of course, if the AI has "memory" or "project" features, this situation might improve somewhat.

But Agents can also, based on tasks you delegate, within authorized information ranges, intelligently make decisions, proactively call multiple tools, then autonomously execute specific complex tasks and deliver the final results to you. You just need to "sit back and enjoy." Moreover, Agents can not only have memory but can also be trained, becoming smarter and smarter through repeated interactions, making results increasingly meet your expectations.

How to Implement Notion Agent

To use Notion Agent, you first need to create a document for this Agent, then define the Agent's basic behavior guidelines in this document, such as identity and mission, communication and behavior guidelines, or work scenarios and action goals, etc.

Assuming your need is to let Notion AI better assist your content creation, the minimum viable document for this Agent could be written like this:

After writing this document, then in Notion AI personalization settings, specify it as the Agent's underlying instructions, as shown below:

This way, before Notion AI gives every answer, it will follow the instructions in this document and interact with you according to your preset behavior guidelines. Thus, facing the same question, with an Agent document set versus without an Agent document set, the answers Notion AI gives will be completely different. The former is more precise and personalized, while the latter is more general but also more mediocre.

But at this point, someone might be curious: Isn't this just giving AI a pre-written prompt in advance? If I manually give this prompt to Doubao or DeepSeek before using them, wouldn't it have the same effect? To some extent, yes, but Notion Agent has several differences from other AI chat tools:

1. Documentation as Rules

Other AI Q&A tools require you to manually input or copy-paste prompts every time, while Notion Agent's rules are written in Notion document pages. They can be automatically loaded, changed at any time, and take effect immediately. More importantly, you can also directly reference other Notion notes you've already written in this Agent document, like this:

That is, you can extremely quickly connect (reference) the creative SOPs you've already written, your recorded writing insights, and your commonly used writing methods into this Agent. Completely without any complex operations, because it is itself a Notion note page.

This creates a feeling of "stepping on one's own feet to ascend to the heavens". Notion itself provides an excellent environment for recording notes, while Agent allows these沉淀下来的经验 to be quickly validated. In your collaboration process with Agent, if you discover the strengths, weaknesses, or gaps in these experience notes, you can further optimize and supplement these notes. A positive feedback loop is established.

This is the meaning of "documentation as rules" - your notes are no longer just static records but executable rules and experiences that can be reused and validated.

1. Agent Can Directly Operate Notion Workspace

This is another underlying capability of Notion Agent, which is permission to perform CRUD operations on pages and databases. Other AI tools can only generate text answers, while Notion AI can directly execute underlying operations. Here are a few specific examples:

1️⃣ Creating and Modifying Notes

When you've finished discussing an idea with AI, you can directly have it organize the dialogue content into a note, then save it to a specified database. AI will automatically read that database's field properties, understand the purpose of each field, then intelligently fill in content: automatically tagging appropriately, relating relevant projects, setting priorities, etc.

For example, I'm currently learning about Claude Skill. I can ask Notion Agent to search the entire web and summarize into notes, then store them in my "Notes Database," as shown below:

Notion Agent not only correctly organized relevant notes but also knew which was my "Notes Database," and even filled in the corresponding database fields for me. It correctly understood the usage of the six basic tags for note types I introduced in this article.

2️⃣ Batch Operations on Database Pages

When you need to batch process tasks, Notion AI can complete multiple operations on a database at once, such as:

• Mark all overdue tasks as high priority
• Find all tasks completed this week and generate a summary

These things that require you to manually click one by one and repeat operations in other tools can be completed by AI with just one sentence in Notion.

Additionally, this example database was actually also created directly by Notion Agent. I told it "Please understand the context information, then create a demonstration database for this sentence."

3️⃣ Workflow Automation

Going further, you can have AI automatically execute complex operation sequences based on specific conditions. For example, the "help me generate a weekly report" task is not simple information query but a complete workflow. This requires accessing multiple data sources → filtering pages that meet conditions → reading page content → generating reports in a specific format → saving to a specified location and automatically filling properties. All steps are still automatically executed according to preset rules without manual intervention.

Later there will be more extensive introduction to automated workflows, so I won't expand here for now.

4️⃣ Real-time Modification of Rules and Memory Based on Requirements

When you first have Agent generate a weekly report and find it summarizes too briefly, you can directly say "Remember, every task in the weekly report must write specifically what was done," and Agent will proactively go modify the rules in the Agent document. Next time it generates a weekly report, it will automatically execute according to this standard. Or you find Agent always over-praises when reviewing articles. You can say "In the future, directly point out problems, don't praise," and it will immediately adjust its response style and automatically modify the Agent document.

Once you're accustomed to this interaction method, updating Agent behavior becomes extremely simple. One sentence can make it remember and execute, without needing to rewrite complex rule documents. Your collaboration will become increasingly tacit.

For example:

The effect is as follows:

The several basic features above together form the foundation of Notion Agent's underlying capabilities:

1. Documentation as rules: Your notes directly become Agent's behavior guidelines
2. Database as memory: Agent knows where to read from and where to write to
3. Dialogue as training: One sentence can make Agent remember and improve

But having underlying capabilities alone isn't enough. How to organize these capabilities and apply them to real work scenarios is what truly tests system design. Next, I'll share some Agent document design ideas to help you connect these underlying capabilities into truly usable workflows.

3. Agent Design Thinking

Scenario Routing

Real work scenarios are complex. When you say "help me take a look," you might be reviewing an article, checking a video script, or reviewing a project's progress. The same words mean completely different needs in different contexts. We could certainly write all situations clearly in the Agent document, but then every time you start a conversation, Agent needs to load all documents simultaneously, which will inevitably waste some context space.

So my suggestion is to use progressive disclosure thinking to build your Agent document, that is, displaying information in layers, diving deeper when needed, avoiding loading all content at once. My personal approach is to first specify that every conversation must load the following four core documents. Every time you start a new conversation, only load the minimum necessary information.

These core documents vary by person. You can flexibly configure them according to your usage scenarios, but generally suggest at least including these elements:

1. Identity and Mission: Define who Agent is, what the core goal is, and who the served object is
2. Interaction Style: Communication tone, response format, when to be concise, when to be detailed
3. Continuously Updated Memory: User preferences, latest habits and requirements, etc.
4. About User: User's identity, background, work style, values, etc.

Then only when Agent recognizes specific keywords in the conversation does it further load subdocuments, just like the diagram below. Each scenario's subdocument contains a complete execution manual defining that process's specific workflow, judgment criteria, and output format. This way, not all subdocuments are loaded at once, making Agent's responses more efficient.

Take this article I'm currently writing as an example. I can directly select article paragraphs, then propose requirements "generate image" in the Agent window. Notion Agent will read the "image generation" keyword, thereby only triggering the "Scenario N: Content Image Generation" subdocument, as shown below:

According to the SOP I predefined in the "Scenario N: Content Image Generation" document, Notion will automatically execute the following operations according to this document's instructions:

1. Select Style: By default, choose the predetermined drawing style
2. Understand Content:
   • If it's a partial illustration: Analyze the meaning of the selected text and the illustration's purpose
   • If it's a full article cover: Extract the article's core theme and emotional tone
3. Generate Drawing Prompt: Based on selected text content + default design style document + context information + additional requirements noted in the chat window, directly output complete drawing prompts

In step 3, I require Notion Agent to prioritize referencing my preset "Top-level Design Style" document. This document specifies the default image generation prompt's style, ratio, preferences, etc. So you can see that the illustration styles generated with Nano Banana in this article are relatively consistent.

The generation process and results are shown below:

The same operational logic applies to other scenarios. For example, I simulated a "Diet Record" subprocess. Then I can directly send food photos to Notion Agent, use the keyword "today's food" to trigger this "Diet Record" SOP, and Notion Agent will automatically analyze the food in the image, then record data like calories, carbs, fats, etc., and save them to a specified database.

It needs to be explained that the calorie data AI evaluates after image recognition isn't completely accurate and can only be used as reference. But the identification of food types in the image is quite simple. So can we use this to build a diet assessment system?

Assume I'm a diabetes patient. I created a "Type 2 Diabetes Personal Health Record" note, then placed this note in the required loading items for the "Diet Record" process, and required Agent to compare the read food one by one with the contraindication list in the record, and explicitly explain in the feedback the food's impact on blood sugar:

After eating, send the photo to Agent and use the keyword "today's food" to trigger this process:

The following is the feedback Agent gave me:

1. Recorded diet data
2. Gave me clear health reminders

Although AI models inherently have the capability to provide reference advice for specific diseases, if personal health records and medical record reminders can be added, the answers' results will inevitably have higher reference value.

From the above cases, we can see that Notion Agent's true power lies in the combination of "keywords + subdocuments."

By triggering keywords to let Agent enter specific scenarios, and the nested subdocuments within scenarios (like health records) further refine execution rules. At the same time, this health record is also a Notion page that can be quickly modified and adjusted. This layered structure allows Agent to maintain generality while showing higher professionalism in specific scenarios.

If you've used Claude's Skill feature, you'll find its logic is similar to "keyword trigger + subdocument execution," both can follow the progressive loading principle.

In comparison, Notion Agent's disadvantage is that the tools it can call are limited to the Notion ecosystem. It can't execute custom script code like Skill can. But Notion Agent only needs to deal with documents throughout the process. The barrier is lower. As long as you can write documents, as long as you clearly express your ideas, even without clearly expressing your needs, you just need to constantly dialogue with Notion Agent, let it ask you questions. Even if your answers are vague, AI models can slowly figure out your intentions, then intelligently help you build the entire process.

Information Boundaries

With scenario routing, we also need to solve another problem: information input sources and information output target locations. I call this problem "information boundaries."

You have Agent generate a weekly report. It needs to know where to read this week's task data. You have Agent save an inspiration. It needs to know which notes database to save it to. If you need to manually specify "read this database" or "save to that location" every time, the usage cost is too high.

Moreover, if information boundaries aren't clear, answer accuracy will inevitably be affected, because Notion Agent's accessible information sources are extremely rich.

But we can't preset in the Agent document "if user asks A, go read page X; if user asks B, go read page Y" in exhaustive detail. This not only has extremely high maintenance costs but also lacks flexibility. So my solution is database structured storage + scenario presets.

Here's a specific example. I have an Agent scenario F for automatically generating daily, weekly, and monthly reports. The trigger keywords are as shown below:

In this scenario F's execution document, I defined the information input sources like this:

This way, when I say "generate this week's weekly report," Agent knows:

• Which databases to query for information
• What the query filter conditions are
• No need to search other irrelevant pages or databases

The query results are as follows:

After querying the specified information, then tell Agent how to process this information:

After this, there will be many rules for analyzing and processing information, but ultimately the core is to directly help me generate a systematic complete report according to the preset weekly and monthly report templates.

These clear boundaries bring three benefits. First, Agent won't aimlessly search the entire workspace but precisely obtain information from specified data sources. Second, clear data sources mean faster query speeds, not wasting time on irrelevant information. Most critically, you know where Agent will read information from. Its behavior becomes predictable and controllable. If a result isn't ideal, you can also quickly locate the problem: whether the data source's information is incomplete, or Agent's extraction logic has issues.

However, all this premises that your Notion workspace itself relies on structured databases:

• Tasks have fixed storage locations
• Notes are classified and stored by type
• Project information has clear organization
• Web-clipped articles have an independent management center

That is, Notion Agent's capabilities must be built upon Notion's own organizational capabilities. If your workspace is a mess, Agent also can't function well. But if you have clear information architecture, Agent can accurately locate needed information and give precise results. Many people try Notion AI and feel it's dispensable and useless, essentially because they lack two things: rich records and clear structure.

Solving where information comes from, next use the same method to solve where information should go.

If after every interaction with Agent, you need to manually specify which database to save to, which fields to fill, what tags to set, then the usage experience will be very poor, and automation of information flow becomes out of the question. So my solution is still preset storage rules + intelligent field filling.

Notion Agent's power lies in that it can not only create pages in databases but also understand database structure and intelligently fill various fields.

Continuing the monthly report example above, the monthly report Agent generates won't stay in the chat window waiting for me to manually copy-paste but will automatically save to the "My Notes DB" database, because I've already noted the following elements in this process's document:

1. Storage location for generated monthly reports
2. Format template for generated monthly reports

After Agent accesses the "Notes Database," it goes to understand the database fields' semantics. It knows that in the "Note Type" field, Exp represents experience review, Idea represents inspiration record, Log represents log record. So this monthly report was automatically tagged with Exp. This semantic understanding capability makes the entire information flow more automated.

Reviewing the two subsections above, "scenario routing" solves the problem of how Agent should think, while "information boundaries" solves the problem of where Agent should look and where it should write. Only by combining the two can Agent have enough intelligence to understand your intentions and enough constraints to avoid mistakes.

Behind these two designs is a shared philosophy: the more automated the system, the clearer the boundaries needed. If you restrict nothing, Agent's behavior becomes unpredictable. You'll never know where it will find information or where it will store results. But if you clearly define boundaries, Agent's behavior becomes controllable and predictable, and problems are easy to troubleshoot.

Of course, these boundaries aren't unchanging. As your work style changes, database structure adjusts, or you discover certain scenarios' rules aren't perfect enough, you can update these rules through dialogue at any time. This "rules are iterative" characteristic gives Notion Agent both the stability of a structured system and retains sufficient flexibility.

Custom Agent

The content in this next part involves Notion's not-yet-officially-launched Custom Agent feature. I also don't currently have this permission, so I can only reference currently disclosed information. But I can roughly explain what this feature is and what effects it can bring.

All the content introduced above, including scenario routing, information boundaries, documentation as rules, is essentially conducted within the "Personal Agent" framework. That is, all operations require you to actively initiate requests for Agent to execute. You need to open Notion, open the AI window, input commands, then wait for results.

But what Custom Agent aims to solve is: can Agent run automatically in the background without me manually triggering it every time?

Imagine these scenarios:

• At 9 AM, Agent automatically scans your task database, organizes tasks due today into a list and pushes them to you
• Friday afternoon, Agent automatically reads this week's task completion status, generates a weekly report, and saves it to a specified location
• Every weekend, automatically searches the entire web for AI hot news, then summarizes into a brief and saves it to the web-clipping database

This is the core value of Custom Agent - evolving from "you ask, it answers" to "it proactively does."

Reviewing the Agent design thinking we introduced earlier, Custom Agent is essentially an extension of the same logic:

• Scenario routing logic still applies, only the trigger method changes from "keyword recognition" to "time or event trigger"
• Information boundaries design becomes even more important, because automatically running Agent must precisely know where to read from and where to write to
• Documentation as rules philosophy remains unchanged. You still define Agent behavior by writing documents

If you're already using Personal Agent and have established comprehensive scenario documents and information structure, then upgrading to Custom Agent in the future will be very smooth: you only need to change originally manually triggered scenarios to automatic triggers.

Each of our work contains大量的regular tasks: daily and weekly reports, data organization, information synchronization, regular reviews... These tasks aren't difficult individually, but precisely because they're not difficult, they're often procrastinated or forgotten. Custom Agent's value lies in turning these "should do" things into "automatically do" things, letting your energy truly focus on work that requires creativity.

Of course, this also places higher requirements on Notion workspace organization. A disorganized workspace, even with Custom Agent, is hard to function effectively. So if you're interested in this feature, you can start organizing your information structure now to prepare for the future.

Agent Design Template Sharing

If you're a complete beginner, the content above might seem a bit scattered and complex, but actually, the core thinking is quite simple.

A basic Agent document framework usually contains four modules: identity and mission, interaction style, scenarios and trigger words, memory area. You don't need to complete all content at the start. Start with one most commonly used scenario, use it first, then slowly expand.

And there's an even more effortless way to start, which is to use your past records to train AI**.**

Many people face a blank Agent document and don't know where to start, don't know how to define their "identity," describe their "style," or organize their "values." But actually, you don't need to write these things from scratch. Throw all the notes, articles, project reviews, or even casually recorded ideas you've written in the past to AI, let it help you analyze and extract, then profile you. It can identify your expression tendencies, areas of expertise, and judgment criteria for certain things from your writing.

This is the thinking of "using existing inventory to start incremental growth." Low barrier, quick start, and more authentic, because this content is originally yours. AI just helps you organize it.

This also highlights a more underlying viewpoint: recording is the infrastructure of the AI era. Only if you had the habit of recording in the past can you quickly provide corpus for AI analysis now. It's never too late to start recording, because you don't know what AI tools will be born in the future. But no matter how advanced the model, it's unlikely to have mind-reading capabilities. You still need to provide it with analysis materials. No input, no output.

To help you start this cycle faster, I designed a Notion Agent Basic Template. You can directly copy and use it. Click me to get template link. You only need to according to the template's structure and guidance, start dialogue with your Agent, then gradually adjust and optimize in actual use, and add new subdocuments, adjust trigger keywords, and supplement your unique methodology and preferences according to your work scenarios.

Of course, if you want Agent's capabilities to reach their ultimate, a structured Notion workspace is the prerequisite. If you haven't established your information architecture yet, or feel confused about how to organize tasks, projects, and notes, you can refer to my FLO.W Template. It presets clear database structures: tasks, projects, notes, and web-clips each have independent storage locations. Field design has also been repeatedly polished and can be directly recognized and invoked by Agent. You don't need to build information architecture from scratch. The template itself is the foundation of Agent capabilities.

Welcome to learn and purchase:

• FLO.W Template Purchase
• Template Full Process Video Introduction
• Ten-Thousand-Word Interpretation of Template Underlying Design Thinking

Notion AI Subscription Advice

What Notion Agent can do has basically been covered in the previous text. However, given that Notion's capability range is truly diverse and strange, I've also compiled a collection of operations Notion Agent cannot do. You can specifically refer to this link. Therefore, before subscribing, you can first refer to this document to evaluate whether your needs can be met.

Additionally, Notion's official payment strategy has already undergone a round of adjustments. Now if you want to use Notion's AI features, you can only purchase the Commercial or Enterprise version. For personal use, the Commercial version is sufficient, but $240 per year is truly not low. Therefore, my suggestion is that if you're a beginner trying Notion for the first time, I completely don't recommend you subscribe to AI immediately. You can first read the "5 Beginner Tips for Notion Newcomers" I wrote, see if you can accept and adapt to Notion's recording method. Wait until you really feel Notion is a handy tool, then consider whether to subscribe.

Some users might also discover that they don't need to subscribe to the Commercial version but can supplementally subscribe to Notion AI through add-on on top of the Plus version. According to Notion's official statement, this policy is a special scheme for historical subscribers. Since May 2025, AI add-on is no longer sold separately to new users. That is, only old users who had already subscribed to AI add-on before the policy adjustment can continue using Notion AI in this way. Moreover, subscribing to AI features this way is incomplete. The main functions lack the following points:

• AI Agent: Personal AI assistant that can execute multi-step tasks
• Enterprise Search: Global search across workspaces and connected applications
• AI Meeting Notes: Automatic voice transcription and meeting summary generation

And it needs special attention that once add-on subscription is canceled, it cannot be restored. At that time, you can only regain complete AI features by upgrading to Commercial or Enterprise versions.

Conclusion

Most people's relationship with AI still remains at the stage of "opening a chat window when there are problems" - use it and leave, next time meeting again as strangers.

But Notion Agent gave me a different possibility: AI is no longer an external tool but a collaborator that can be trained, shaped, and grow together with you. It won't ask "what format would you like" every time, because it already remembers your preferred way. It won't aimlessly search in your workspace, because you've already told it where to look and where to store. This feeling is wonderful, like cultivating an assistant, an assistant who becomes increasingly tacit over time.

Of course, all this premises that you're willing to invest time to build, train, and iterate this system, and more importantly, get your hands dirty and record honestly. Notion Agent isn't out-of-the-box magic. It requires you to first think clearly about your work style, then express it in document form. This process itself is a form of self-organization. You'll discover many habits and preferences you were never aware of before.

This also leads to another viewpoint I want to express: the current AI model capability gap is shrinking. What truly determines output quality is the input you feed it.

But the "input" here doesn't refer to those ubiquitous, already flooded second-hand information, not asking you to web-clip others' articles or collect others' views. Even worse, these information might actually be excrement processed by AI. Truly valuable input is your own things: your biases, your ignorance, your narrow perspectives, your clumsy operation process, the detours you've taken, the pits you've stepped on.

Only when you honestly record these "imperfections" can AI truly help you. Because what it sees is no longer cookie-cutter correct answers but your unique thinking trajectory. It can understand your true confusion from your mistakes, identify your true needs from your preferences, and learn your true standards from your repeated revision process.

So I'm quite opposed to those voices promoting "note-taking useless theory in the AI era". In my view, this is just a statement of giving up in the face of the AI wave impact, an excuse for laziness. Any era's, any tool's changes cannot shake the importance of independent thinking. AI can help you execute, help you organize, help you accelerate, but it can never replace you thinking about "what do I really want."

The greatest value of Notion Agent to me is that when I saw the huge potential of automated workflows and the clear implementation path, I realized that I really couldn't delay any further. I must seriously think about what my workflow really is, what mechanical, repetitive processes are consuming my time, and how to adjust and optimize to complete more things in limited time. Once thought clearly, Notion Agent should play its role.

Finally, this article was originally planning to compare with Obsidian's AI plugin and Heptabase's iteratively improved AI, and also hoped to deeply discuss Notion AI's current functional limitations. But alas, the length is already approaching ten thousand words, so this content will be left for the next article.

If there's any content you hope to see, you're also welcome to leave comments in the comment section. I'll evaluate and include it in the writing scope.

---