Switching my second brain: Obsidian, Claude Code, and Karpathy's idea

I've been using Notion as my second brain for a few years. It works, mostly. But it's never felt like something that thinks with me. It's more like a very organized pile of notes I occasionally dig through.

Then I came across something that made me rethink the whole setup.

Karpathy's approach

Andrej Karpathy, former OpenAI, now doing his own thing, posted about how he maintains a personal knowledge base of around 100 articles and roughly 400k words. The idea is simple: once the knowledge base is large enough, he can ask his LLM agent complex questions. The agent researches against the wiki, synthesizes answers, and auto-maintains index files and brief summaries so it always knows what's in there.

No vector database. No embedding pipelines. No RAG setup. Just a well-organized wiki and an LLM that can read it.

That's the part that struck me. We've been overcomplicating this.

The stack that makes it work

The setup has three layers, and each one does a distinct job:

Obsidian (the knowledge layer). If you haven't used it: it's a local, offline notes app where everything is plain markdown files. Think Notion but owned by you, no cloud dependency, no subscription required, infinitely linkable. Your projects, meeting notes, research, thoughts, all in one place, all searchable, all yours, all free. The fact that it's plain markdown is the key thing. Any tool can read it.

Claude Skills (the instruction layer). There's a repository by one of the engineers at Obsidian (kepano/obsidian-skills) that gives Claude Code context on how to work with Obsidian specifically: how to navigate markdown, use the CLI, work with Bases and JSON Canvas. This is the glue between the knowledge and the thinking.

Claude Code (the thinking layer). Your actual AI agent that reasons over everything. With access to the Obsidian vault and the right skills loaded, it can answer questions like "what did I learn about GTM strategy across all my projects" and go find the answer in your notes.

The web clipper bonus

There's a web clipper extension that saves any webpage as a markdown file straight into Obsidian. Anything interesting you find online becomes part of your knowledge base instantly. It struggles a bit with images, but there's a plugin for that. One thing worth setting up: make clipped notes land in a /raw folder so your vault stays organized.

Why I'm switching

The Notion setup works but the knowledge just sits there. It's searchable but it doesn't get smarter over time, and I can't really ask it things. This stack makes the knowledge base queryable in a way that feels closer to how I actually think. Not searching for a specific note, but asking a question and getting a synthesized answer from everything I've read and worked on.

That's the version I want: everything I've read, learned, built, and thought about, something I can ask questions to. That's genuinely useful.

Worth exploring if you're thinking about better knowledge management. The full deep-dive gist from Karpathy is the best starting point.

---

Possibilities this opens up, both personal and professional:

• Ask "what patterns have I seen across all my client projects" and get a real answer
• Give an AI agent your entire meeting history and ask it to identify recurring themes
• Build a research layer for a specific domain (competitive landscape, tech trends, a niche market) that you can query conversationally
• For businesses: a queryable knowledge base for onboarding, institutional memory, and product decisions
• For solo operators: a system that compounds everything you learn instead of letting it decay in folders