Ben Stegink
A concept quietly gaining traction in the world of AI offers a genuinely practical way to manage personal knowledge: the LLM wiki. The idea is straightforward, the setup is lightweight, and the results scale surprisingly well into professional workflows, including those centered on Microsoft 365.
What an LLM Wiki Actually Is
The core concept, popularized by Andrej Karpathy in a public gist, treats a folder of Markdown files as both a knowledge base and a processing target. You accumulate raw source material — articles, documentation, clipped web pages, meeting notes — in a designated intake directory. An LLM then reasons over that material and extracts structured outputs: entities, concepts, source summaries, and inter-document relationships. The result is a self-organizing knowledge graph you actually control.
Markdown is the right format for this. It is plain text, universally readable, and something LLMs handle with particular fluency. It also makes the entire vault portable across tools, cloud storage providers, and AI interfaces.
Obsidian works exceptionally well as the vault layer. Its wiki-style internal links, tagging system, and knowledge graph visualization give you a rich interface on top of what is ultimately just a folder of text files. Sync it through OneDrive or Google Drive and it follows you across devices automatically.
Getting Started
The fastest path to a working setup is to take Karpathy’s gist, paste it directly into your preferred AI interface (Claude Code, GitHub Copilot CLI, Gemini CLI, or similar), and append a single line of context: something like “create a second brain for me that manages everything I want to read and research.” That one prompt generates a context file, typically a claude.md or copilot-instructions.md, that defines the vault’s structure, ingest rules, and output templates.
That context file is the durable part of the system. It is portable across AI interfaces, meaning you can use the same instructions whether you are running Claude locally, via the API, or through GitHub Copilot. If you want a pre-built starting point rather than generating from scratch, this GitHub Copilot LLM wiki template provides a forkable structure with the directory layout and templates already in place.
Once you have a context file, iteration is fast. Feed a handful of articles into your intake folder, run an ingest command, and watch the wiki populate. Entities link to concepts, concepts link to sources, and Obsidian’s graph view starts to show you relationships you did not explicitly create. If you do want to use this with Microsoft 365 Copilot Cowork, ensure that you create your folders and files within the Cowork folder in your OneDrive\Documents location. This way, Cowork will have the ability to both read and write files within your wiki.
Keeping Token Costs Under Control
The practical wall you may hit quickly is token consumption. Ingesting even moderate amounts of content gets expensive fast, especially if the LLM is re-enumerating your directory structure and re-reading template files on every run. The solution is to think like a developer from the start: extract the repetitive operations into scripts.
After a few ingestion cycles, ask your LLM to analyze what it keeps doing repeatedly and automate those steps. A simple Python or PowerShell script handling directory scanning, file comparison, and template validation can eliminate a significant portion of token overhead before the model ever sees your content. This llm-wiki-manager project takes a similar approach to systematizing wiki management. The context file evolves alongside the scripts, replacing verbose procedural instructions with lean references to pre-built commands.
For a no-code starting point, this walkthrough using Claude or GitHub Copilot demonstrates how to stand up a working wiki without writing any scripts yourself, which is useful for initial experimentation before you start optimizing.
Connecting It to Microsoft 365
Microsoft Copilot Cowork adds a compelling layer for anyone living inside Microsoft 365 (and if you are weighing it against Anthropic’s desktop-first alternative, we compared the two in Claude Cowork vs. Copilot Cowork). Rather than manually clipping content, you can build skills that extract email from your inbox, convert each message to Markdown, archive the originals, and deposit the results directly into your ingest folder. Run that skill on a schedule and you effectively stop living in Outlook — the content you would otherwise triage by hand flows into your second brain automatically, where the LLM organizes it alongside everything else you are researching.
That is the real payoff of building on Markdown and a portable context file. The same vault that started as a folder of clipped articles becomes a knowledge base your AI tools can reason over from any direction: locally through Claude or Copilot CLI, visually through Obsidian, and automatically through scheduled Cowork skills tied to your Microsoft 365 data.
Conclusion
An LLM wiki is one of the rare productivity systems that gets more valuable the less you fuss with it. Start small with Karpathy’s gist and a single context file, let scripts absorb the repetitive work to keep costs in check, and then wire in Copilot Cowork so your Microsoft 365 content feeds itself into the vault. Just go in aware of Cowork’s current security blind spots if work content will be landing in your vault. If you want help connecting a setup like this to your tenant — or building the Cowork skills to automate the ingest — reach out and let’s talk.
