LLM Wiki: Building a Self-Evolving Obsidian Second Brain with AI Agents
中文What
LLM Wiki is an AI-orchestrated knowledge management workflow. In this workflow, humans provide raw materials (articles, books, PDFs), while AI agents autonomously organize, format, cross-reference, and index them, ultimately integrating them into a highly structured and bi-directionally linked Obsidian vault.
How
1. Build an Obsidian vault containing designated folders: 'Sources' (for raw materials), 'Knowledge Base' (for structured classifications like people, events, etc.), 'Attachments', and two root directory files: 'Index' (Map of Content/MOC) and 'Log' (history logger). 2. Create an AI instruction file at the vault's root (named 'CLAUDE' for Claude Code, or 'AGENTS' for Codex) detailing folder structures, Markdown templates, and processing rules. 3. Connect a local agent tool (such as OpenAI Codex) to the vault folder. 4. Direct the agent to import files from 'Sources', format them using templates, place them in 'Knowledge Base', generate bi-directional links, and update the 'Index' and 'Log' files. 5. Set up automated routine tasks (e.g., checking the source folder every day at 9 AM, maintaining broken links once a week) and use the Obsidian Web Clipper to quickly import web articles into the system.
Why
Over time, manually organizing vaults, applying templates, and finding bi-directional links can become tedious and unsustainable. Delegating file classification, cross-referencing, and indexing to AI agents retains the utility of a highly interconnected 'second brain' while removing the hassle of manual maintenance, allowing users to focus on reading and synthesized learning.
Other
This workflow relies heavily on the AI agent having local directory read/write permissions. Clear prompt engineering in the system's instruction files is crucial to prevent the AI from damaging vault templates, corrupting folder paths, or creating duplicate entries.