I Built Karpathy's AI Second Brain With Obsidian and Claude Code: Here's Exactly How

Every second brain system ever sold has the same failure buried in it: you are the librarian. You capture, you file, you link, you maintain. The system works exactly as long as your enthusiasm does, which for most of us is somewhere between eleven days and three weeks. Then the inbox fills, the guilt accumulates, and you are back in Apple Notes.

In April this year, Andrej Karpathy, the former Tesla AI director who coined the term vibe coding, published a short GitHub gist that flips the whole premise. He calls it the LLM Wiki, and the idea fits in one sentence: your second brain should not be written by you. It should be written by an LLM.

You drop sources into a folder. A coding agent reads them and compiles interconnected wiki pages. You ask questions and get answers drawn from your own curated material. You are no longer the librarian. You are the collector and the reader, which are the two jobs you were always going to actually do.

I spent two weeks building and living with this pattern, using Obsidian as the viewing layer and Claude Code as the agent. Here is the honest walkthrough.

Why this pattern is different, mechanically

The traditional second brain fails at the filing step. The LLM Wiki removes the filing step by making one architectural choice that matters more than any app feature: everything is plain markdown files in a folder.

That choice is why Obsidian is the right viewer and not Notion or Evernote. An LLM agent can read, write, link, and reorganise plain text files directly. Proprietary databases require APIs, permissions, and hope. A folder of markdown requires nothing. Your notes become a workspace the AI can maintain, and simultaneously remain files you own forever, readable by any text editor on earth.

The build, step by step

Step 1: The vault and the two folders.

Install Obsidian, create a vault, and inside it create exactly two folders: inbox and wiki. My first attempt had seven folders with a taxonomy I was proud of. Delete your version of that. The entire point is that structure is no longer your job.

Step 2: The Web Clipper.

Install the Obsidian Web Clipper browser extension. It strips any article down to clean markdown and saves it into your vault locally, no cloud middleman. Point its default save location at inbox. From this moment, capturing anything from the web is one click, and it lands as raw material the agent can process.

Step 3: The agent instructions.

Open Claude Code in your vault folder and create a file called CLAUDE.md. This is the librarian’s job description. Mine, after two weeks of refinement, says roughly:

Process everything in inbox. For each item: identify the core topics. For each topic, create or update a page in wiki that synthesises what my collected sources say about it, written concisely, with links between related wiki pages using Obsidian’s double-bracket format. Every claim in a wiki page must trace back to a source file. Move processed items to inbox/processed. Never delete anything. Never invent information not present in the sources.

That last instruction is not decoration. An agent left without it will helpfully fill gaps with its own knowledge, and the entire value of this system is that it answers from your material, not the model’s general opinions.

Step 4: The first run.

I seeded the inbox with about thirty clipped articles and ran Claude Code with one instruction: process the inbox per CLAUDE.md. Eight minutes later the wiki folder contained fourteen interlinked pages, each synthesising multiple sources, each linking to its neighbours.

Opening Obsidian’s graph view at that moment is the advertisement for the whole pattern. A knowledge network existed, and I had not filed a single thing.

Step 5: Living with it.

The ongoing loop: clip things all week with zero thought. Once or twice a week, run the agent to process the inbox. Ask questions directly in Claude Code when you want answers: “what do my sources say about X” returns a synthesis with the wiki pages as its footing.

Where it got weird, because it did

Honesty section. Three real problems from two weeks of use.

The agent has editorial opinions. Left unwatched, my wiki pages drifted toward a tone I would call enthusiastic consultant. Adding “write plainly, no marketing language” to CLAUDE.md mostly fixed it. Mostly.

Duplicate concepts creep in. The agent created both a “Prompt Engineering” and a “Prompting Techniques” page before I added an instruction to check for existing pages first. Plan to garden lightly in the first week while the structure settles.

Trust requires spot-checking. Twice in fourteen days a wiki page contained a claim that was a reasonable inference rather than something my sources said. The traceability rule in CLAUDE.md makes this checkable, and you should actually check, especially early. The system earns trust the way any librarian does: gradually, with audits.

The verdict after two weeks

Every previous second brain I built died of maintenance debt. This one has no maintenance debt, because maintenance is the machine’s job and the machine does not get tired of it. Fourteen days in, clipping is still one click, processing is still one command, and for the first time the graph view shows a structure I did not build and do not resent.

The pattern will not suit everyone. It requires running a coding agent, and that step, while genuinely not coding, is a real threshold. But if you are the person with three abandoned vaults behind you, this is the first version of the idea where the part you kept abandoning is no longer yours to abandon.

Karpathy’s original gist is public on GitHub, and community implementations are already multiplying around it. The version described here needs nothing beyond Obsidian, the free Web Clipper, and a Claude subscription you may already have.