This series is about how agentic AI can be used to innovate the way we work with our information and knowledge (aka knowledge work). While using LLM (e.g. ChatGPT) helps with certain aspect of knowledge work such as essay proofing or brainstorming, agentic AI has potential to dramatically improve our productivity by integrating AI deep into our workflow.
Where do we start? While we don't have generally available AI agent for knowledge work just yet[^1], we've found that combining filesystem-based coding agent (e.g. Claude Code) with filesystem-based PKM[^2] solution (Obsidian) gives us pretty decent agentic AI experience for knowledge work. In this series, we'll share what we learned along the way, and how you can get started on your journey.
[^1]: Notion AI Agent can be an early attempt, but it's restricted to Notion users. [^2]: PKM means Personal Knowledge Management, which can be interpreted as document management software for now.
💡 Presentation Version: This introduction can be adapted into slides for presentations. The key visual elements include the dial-up modem analogy, workspace integration diagrams, and the agentic AI architecture overview.
p.s. For more deep dive onto PKM aspect of this project, please refer to AI for PKM series.
Why ChatGPT is not enough
The limitations of standalone AI tools become apparent in real-world knowledge work.
%%Can we bring AI into our workspace?
The problem with current AI workflow is that user's workspace is separate from AI service, and the communication is limited in bandwidth (e.g. output buffer size) and manual (attaching context file; pasting AI output back to the doc)
Human-AI collaboration model: (limited bandwidth)
%%🖋 Edit in Excalidraw%%
This workflow is fine for simple work, but doesn't scale well to most real-world problems. This almost feels like the first generation of internet where we used dial-up modem to connect websites. In order to dramatically improve the communication bandwidth, can we bring AI into our workspace? (e.g. file system)
Human-AI collaboration model: (shared workspace)
%%🖋 Edit in Excalidraw%%
Agentic AI Solution for Knowledge Work
Our proposed solution looks as follows 1. Filesystem-based PKM (e.g. Obsidian) for human user 2. Filesystem-based agentic AI (e.g. Claude Code or Cursor) as AI layer 3. Collaborative human-AI workflow for efficient supervision and feedback
%%How Agentic AI Changes the Game
| Changes | Descriptions |
|---|---|
| Better Input (Context) Management | AI Agent can find context documents automatically from user's file system |
| Increased Output Bandwidth | AI Agent can directly create or modify multiple documents. |
| Better Prompt Management | A collection of prompts can be stored and reused by the system |
| From Prompt to Workflow | A prompt can be combined to form a more complex workflow |
Wait, but aren't these coding tools? - Yes, they are, but coding is a specific type of knowledge work, where input/output is code - We argue that the capability of coding AI agent is generalizable beyond the coding domain - It can efficiently read and write multiple documents - It can use various file management tools including search - Working with document is natural for LLM which is originally trained to predict human language
How is this different from NotebookLM?
NotebookLM provides a more convenient way to provide your documents as context to AI, which is a step in the right direction. However, it still has limitations:
- Manual document selection: You still need to manually select and upload documents for each query
- Output copying: You still need to paste AI outputs back into your target documents
- Limited integration: It operates as a separate service rather than being integrated into your workspace
Our proposed agentic AI solution addresses these limitations by bringing AI directly into your file system, eliminating the manual steps and creating a seamless workflow.
In the following parts, we'll explore the individual components of this proposed workflow and examine detailed examples of AI-supported personal knowledge management and content creation processes.