Second Brain Starter for Claude Code: Karpathy's LLM Wiki

Context

In April 2026, Andrej Karpathy, a co-founder of OpenAI, published a short note called LLM Wiki. You collect raw material: notes, articles, documents. An AI reads it and writes a wiki from it, a set of pages linked to each other. You almost never write those pages yourself. The AI keeps them up to date.

I built my own system on this idea first, and I wrote about it in an earlier article. Over the months it grew into a control centre for my work. It knows how I work, and it is built for exactly one person.

Then I wanted to give something like it to other people.

Problem

My own system was the wrong thing to hand over. What makes it useful to me is everything I put into it: the knowledge, the commands, the way of working. Someone else would not use it, because it is mine. To them, a copy would be someone else's memory.

A ready-made template has the opposite problem. It gives you a folder for every topic it expects you to have. A week later, most of those folders are still empty.

And the people I had in mind do not write code. Many of them have never opened a terminal, the window where programmers type commands. If the first ten minutes feel technical, they stop.

What it had to get right

Set up without a terminal, just by talking.
Grow from the user's own work, not from a ready-made template.
Never sound more sure than it is. A guess has to look like a guess.

What I built

Second Brain Starter is a folder that becomes a memory you can talk to. It runs in the Claude desktop app, in the tab called Code. You do not write any code there. You just talk. The Code tab needs a paid Claude plan. There is nothing else to install.

You make an empty folder, open it in the Code tab and paste one ready-made message from the starter's GitHub page. Claude downloads the starter into the folder and starts the setup. From there, everything happens in conversation, in whatever language you write in.

Setup starts with one question: your name. Then it asks for one real thing to save: a document or a note you are working on this week.
It asks before it saves. Hand it a file or paste some text. Before it writes its pages, it asks what it needs to know: what this is, who it is for, why it matters to you.
It writes the pages. Your file stays as it is, in a folder called raw. The AI writes its pages in a folder called wiki and links them to what is already there. Every page says how sure it is.
Ask it later. A week later, ask what you agreed with a client, or what is still open. It answers from your own pages and links the ones it used.
You tell it when you are done. Decisions you made in the chat get written into your pages before you leave. If the chat winds down and something is still not written, it offers to do it.
Finished work gets its own folder. Ask for a summary or a draft reply, and it lands in a folder called out. It stays apart from the pages it was made from.

Second Brain Starter demo in the Claude desktop app, invented example: the owner drops in a builder's kitchen quote and writes 'this is for the base'. Claude says it saved a copy, sums up the quote, points out that it has no price, that appliances and painting are not included and that the deposit is due before 3 November, lists the three pages it plans to write, and asks which part matters most before it goes ahead. — A demo with an invented kitchen job. The first file is a builder's quote. Before it writes any pages, it says what the quote is, points out that it has no price, lists the pages it plans and asks what matters to you.

If you know Karpathy's note, all of it is here. His three layers: raw/ for the sources, wiki/ for the pages the AI writes, and one rules file, CLAUDE.md. His three operations: ingest a file, ask a question, and lint, which checks the wiki for gaps and contradictions. I added two things his note does not have: out/ for finished pieces, and a wrap-up when you say you are done.

The pages are simple text files in ordinary folders. They stay on your own computer. Claude sees what it reads from them, the same as anything you type into a chat. If you have Git, a free tool, it saves every version in the background. Nothing gets lost, and you never have to learn Git.

Obsidian, a free app, shows the same pages as a map of linked notes. When I set it up with someone, I open Obsidian on day one. Seeing the pages connect is when it starts to make sense.

Second Brain Starter demo base in Obsidian's graph view, invented example: six dots joined by lines, one per page the AI wrote from three files (a builder's quote, notes from his visit, his email): kitchen-renovation in amber in the middle, dan-novak, leo, and one page for each of the three files. — Each dot is a page the AI wrote, each line a link. The same demo after three files: the builder's quote, notes from his visit and his email. The amber dot is the main page, the kitchen job itself.

Guesses stay guesses

An AI that writes your pages will sometimes get something wrong. So the base never hides how sure it is.

Every page carries a mark at the top: seed for a guess, working for partly checked, settled for checked. The AI suggests the mark. Only you can mark a page as checked, and you say what you checked it against.

Every claim links to the file it came from. When new material disagrees with a page, the AI does not overwrite it. It puts both versions on the page and tells you. When you correct it, the fix goes into the page, not only into the chat. If an answer rests on unchecked pages, it says so. Its own drafts never count as sources, so it does not quote itself as fact. A check you run every few weeks finds contradictions and asks about guesses that never moved.

Second Brain Starter demo page in Obsidian, invented example: a Contradiction box keeps the start date from the builder's quote, 3 November, next to the one from his later email, 24 November, each linked to its source; below it a Clash box works out that four weeks from 24 November ends 22 December, two days after the family arrives for Christmas. — The same demo base. The builder's email moved the start from 3 to 24 November. The page keeps both dates, links each one to where it came from, and works out what the change costs: the job now ends two days after the family arrives for Christmas. The choice stays with the owner.

Result

In September I set it up with someone on a call, on their own laptop. By the end of the call it was running, and it already knew their first deal.

The first file was a property offer, a PDF. Before it saved anything, it asked two questions: what is your role in this deal, and who is the client? It did not guess who was who. Then it turned the offer into pages and linked them together.

That was the moment they understood it: the AI wanted to know how they would use their files.

The call also showed a gap. The install steps did not say that you need a paid Claude plan. They do now.

The starter is public and free, under the open MIT license. Anyone can download it and set it up alone.

It starts almost empty, on purpose

The starter gives you the start, not the result. Everyone gets the same base: two empty folders, one for your files and one for the AI's pages, a few commands, and the same rules for what counts as checked. What fills it is your own work.

A ready-made template, a week later

Projectsempty
Areasempty
Resourcesempty
Meetings1 note
Peopleempty
Archiveempty

The starter, a week later

raw3 files you gave it
- 2026-10-01-quote-from-builder
- 2026-10-05-notes-from-dans-visit
- 2026-10-07-email-from-dan
wiki6 pages it wrote
- kitchen-renovation
- dan-novak
- leo
- 2026-10-01-kitchen-quote-from-dan
- 2026-10-05-dans-measuring-visit
- 2026-10-07-dan-email-start-date-moved

A typical template has a folder for every topic it expects, most still empty. The starter, after the three kitchen files from the demo above, holds only what your own work put there.

When you keep asking for the same thing, the AI can turn it into a new command. Next time, one word is enough. Your own commands are listed apart from the ones that ship with the starter. You can always see which part is shared and which part is yours.

The hard part

The hard part was testing the setup the way other people would use it.

I tested it in two ways. First I went through the setup myself, three times. Then I had Claude go through it many more times, each time playing a different user: someone who has never seen a terminal, someone who types commands every day, someone who already keeps notes in Obsidian. Each of them broke something different. The beginner was handed English command words to learn by heart, while the person who types commands every day was talked to like a beginner. The Obsidian user was introduced to an app they already open every day.

The tests also changed how it talks. It now reads your technical level from the way you write. A beginner never hears the name of a command; they just say what they want. Someone who already writes in commands gets the full list.

The first real setup showed the biggest problem. Before you could give it anything, the setup asked five questions about you. It was too much. The same day, I cut them down to one: your name.

I had also pitched that first version as a way to stop losing things. That person did not feel that problem, so I dropped that pitch. Now, right after your name, the setup asks for one real thing from this week.

From the first version, the setup had one goal. You leave with one real thing saved, not an empty folder. Every fix served that goal: nothing in the first hour should make you want to stop. The idea is easy to explain. The setup is the hard step, so I also offer to do it with people on a call.

What this actually says

I built a control centre for myself and found I could not give it away. What I could give away was the start: the same structure and the same rules for everyone. What grows on top of it has to come from the person who uses it.

A memory you can talk to. Made for people who don't code.

At a glance