How to build second Ai-Brain With Claude and Obsidian

Karpathy built a wiki that writes itself. Here is exactly how to copy it.

Most people use Claude like a vending machine. Drop a question in. Get an answer out. Close the tab. The machine forgets everything. Tomorrow you feed it the same coins again.

A 34-year-old consultant in Amsterdam ran this loop for a year. 300+ Claude conversations. Notes in 4 different apps. Browser tabs he kept open for weeks because closing them felt like losing the thought. Every Monday he spent 25 minutes rebuilding context he had already built the Monday before.

Then he stopped pasting context into Claude. He started pointing Claude at a folder.

6 weeks later he opened his Obsidian graph and froze. The system had written 280 connected pages he never typed. It knew things he had forgotten he knew.

Here is that build, written for someone who has never opened either tool.

What you are actually making

Not a notes app. A brain that compounds.

The difference matters. A notes app stores what you put in. This system takes what you put in, traces its connections across everything you have ever saved, and builds a map of your thinking that gets sharper every single day.

2 folders. 1 file. That is the entire structure:

→ raw/ is where you drop sources. Articles, transcripts, PDFs, screenshots. Claude reads these but never changes them. Your sources stay exactly as they are.

→ wiki/ is where Claude writes. Summaries, concept pages, cross-references, connections between ideas. You never write here. The model does.

→ CLAUDE.md is your profile. Who you are, what you are working on, how you want to be spoken to. Claude reads this at the start of every session and never asks you to explain yourself again.

You own raw/. Claude owns wiki/. The moment you blur that line, you lose the thing that makes the system trustworthy.

1. Install 2 free tools

Obsidian from obsidian.md. Free. Stores everything as plain text files on your own machine. Not in a company's cloud. Not in a proprietary format. Just folders of text files that any AI can read, forever.

Claude Desktop from claude.com/download. The desktop version of Claude that can read and write files on your computer. You need the Pro plan at $20 a month. That is the one cost in this entire build. Everything else is free.

2. Create the vault in 2 minutes

Open Obsidian. Click Create new vault. Name it brain. Pick any folder on your computer.

Inside that folder, make 2 subfolders: raw and wiki. Create a blank file called CLAUDE.md at the top level.

That is your second brain. Empty, but built correctly. Everything below fills it.

3. Open the door between the 2 tools

Claude needs a way to reach inside your Obsidian vault. You open that door with a plugin.

In Obsidian: click the gear icon at the bottom left → Community plugins → turn them on → Browse → search for Local REST API → Install → Enable.

Click on that plugin to open its settings. You will see a long string of letters and numbers called an API key. Copy it. Keep Obsidian open. The door only works while the app is running.

4. Connect Claude to your vault

In Claude Desktop, click the Code tab. Paste this line, but replace the placeholder with the key you just copied:

claude mcp add-json obsidian-vault with your key from the plugin settings

The plugin shows the key with the word Bearer in front of it. Do not include that word. Paste only the letters and numbers after it.

To test it, type: list every file in my vault.

If Claude reads your notes back to you, the door is open. If nothing comes back, make sure Obsidian is still running and the plugin is enabled.

5. Load yourself into the brain

Right now Claude can see inside your vault but knows nothing about you. Fix that first, and do not type it out by hand. Make Claude interview you instead.

Paste this into Claude:

You are setting up my second brain. Interview me one question at a time. Ask about who I am and what I do, my goals this year, how I want you to speak to me, my current projects, and my biggest strengths and blind spots. Wait for my answer before asking the next question. When we finish, write everything into CLAUDE.md at the vault root with clear headers.

Answer like you are briefing a new business partner. The more honest your answers, the more useful every future session becomes. When Claude finishes writing, open CLAUDE.md in your vault. That file is now your permanent context. You never explain yourself again.

6. Teach Claude how the vault works

Open CLAUDE.md and add these rules at the top:

raw/ holds my sources. Never edit anything in raw/. wiki/ holds the pages you compile from those sources. I never write in wiki/. When I drop a new file into raw/, read it, write or update the matching wiki pages, and link them to related notes using double brackets. After every ingest, check for broken links, contradictions between notes, and pages that exist but nothing links to.

Now drop a single article into raw/ and say: ingest the new file in raw.

Watch what happens. Claude reads the source, writes wiki pages you did not ask for, links them to knowledge you already have, and tells you about contradictions it found. That is the compile step. That is what separates this from every notes app you have tried before.

7. Run the daily briefing

Once your vault has a few weeks of material, this becomes the most valuable thing in the system.

Each morning, say: read my vault. Tell me what I was working on, what decisions I made last week, what is unresolved, and what I should focus on today.

Claude reads your own notes and gives you a briefing built entirely from your own thinking. Not generic advice from the internet. Your own context, organized and surfaced.

Most people never get this because their notes live in 6 different places and none of them talk to each other. Yours are in one folder. Claude can see all of it.

8. Schedule the vault to maintain itself

This is the step that turns a useful tool into a system that runs without you.

In Claude Desktop, open the Schedule tab. Add a task set to run every morning at 7am. Tell it: check the vault, file anything new sitting in raw, update wiki pages that are now outdated, and write me a 3-line summary of what changed overnight.

You wake up. The brain filed itself. 3 lines waiting for you before you open a single app.

One rule that never breaks: do not rely on words to control what Claude can touch. Saying do not delete this is a suggestion. A read-only permission setting is a lock. If Claude can delete a file, assume one day it will. Control access at the technical level, not the conversational one.

9. Add the things that change daily

A vault of static notes is half a brain. The other half is what changes in real time.

Connect Google Calendar and Claude can log what you committed to in every meeting, flag anything without a clear next action, and surface deadlines you forgot you agreed to.

Connect Gmail and Claude can track the threads that need replies, notice when a client has gone quiet, and file important emails as notes in the right project folder.

Always start with read access only. The brain should read your data and think about it. Not rewrite it.

10. Skip the setup entirely with a ready-made vault

If you would rather clone something that already works than build from scratch, 3 open source repos on GitHub do exactly that.

AgriciDaniel/claude-obsidian sets up the entire Karpathy structure for you, with 15 built-in workflows for different types of work.

eugeniughelbur/obsidian-second-brain ships with 43 ready commands. Save a note, log a daily entry, find something you wrote months ago. Works across Claude, Codex, and Gemini so you are never locked to one tool.

coleam00/second-brain-starter interviews you first, builds a plan based on your answers, and scaffolds the entire vault from that plan.

All 3 are plain text files and scripts. Clone one. Break it. Reshape it into yours.

What you have after these steps

Before: Claude forgets you the moment you close the tab. Every session starts from zero. You are the one holding all the context, and you lose most of it by Friday.

After: a vault that holds your entire thinking. A model that opens already knowing your projects, your decisions, your open questions. A wiki that writes and links itself every morning before you sit down.

Andrej Karpathy ran this pattern on a single topic and watched it grow to 100 sources and 400,000 connected pages without writing a single one himself.

Most people will read all 10 steps and open a new browser tab instead.

The ones who spend 45 minutes on this tonight will open their Obsidian graph in a month and freeze — because the system knows connections they had forgotten they made.

They will never go back to a blank chat box again.