I talk to an AI most days, and for months one thing nagged at me. It is plainly intelligent — it writes, it reasons, it catches what I would have missed. And every morning it wakes up knowing nothing about me. Whatever I explained yesterday is gone. I keep re-introducing my work, my projects, the people around me: a brilliant colleague with no memory of ever having met me.
The obvious fix is to hand it everything — your files, your notes, your saved articles — and let it sort them out. I tried that for months. A pile of documents is not knowledge. The assistant wades through it, answers a little differently each time, and nothing accumulates. The problem was never how much I could feed it. It was that no one was doing the work in between.
That work has a name, though we rarely do it: you read something, understand it, write down what it means, and connect it to what you already knew. It is what a good researcher does. It is also what everyone abandons, because it is slow and there is never an evening for it. So the knowledge stays a pile, and the assistant keeps starting from zero.
Andrej Karpathy sketched a way out in April 2026 — a few pages of notes and a folder layout. What if the machine did that in-between work itself? Not answering on demand, but sitting in the background keeping an understanding of what you know: reading what you give it, summarising it, noticing that today's article argues with one you saved in March and writing that down, keeping the whole thing tidy without being asked.
That is a second brain. Not a notes app and not a chatbot, but a wiki your AI keeps — as much for itself as for you — that grows denser every time you feed it. Mine holds 232 pages now: the things I have read, the ideas that kept recurring, the people and companies behind them, joined by more than nine thousand links. I wrote none of them. I dropped things in and talked about them; the machine did the rest.
Two things still surprise me. It works without me — I need not be at my desk for it to read a transcript and file it. And it finds connections I could not: a line from a podcast that answers a question a book raised months earlier. No person holds that much at once; the machine does, and cheaply, because the filing that bores us does not bore it.
The deeper point is what it does to the assistant. A second brain is a companion tool. It hands your AI what you know in a shape built for a machine to use — digested, structured, cross-referenced — instead of the raw human sprawl it usually has to fight through. The same model, given one, is simply better at helping you. That is what makes it a new primitive: not a place to keep notes, but a memory that makes the intelligence you already reach for sharper.
Seen as memory, the uses spread out. Keep one for your reading, another for your work, one for a single project. A company can build one that amounts to a digital copy of itself — what it is making, what it has decided, what each customer needs — and point its agents at that, so they work on their own because they finally know where they stand. At my scale it is quieter: I research, learn faster, and actually keep what I learn.
The mechanism is almost dull — three folders and a loop with four steps.
Drop. Anything worth keeping goes into a folder called raw/ — an article, a transcript, a note you typed on the train. Nothing in there is ever changed. It is your shelf of originals.
File. You say file it, and the assistant reads the whole thing and tells you what it found, the way a person would. Then it writes a single summary page — where this sits against everything already there, not a re-run of the text.
Connect. Recurring ideas get pages of their own, and so do the people and companies that keep appearing. Every page links to the ones it touches, which is what turns the collection into a web instead of a stack.
Compound. When a new source contradicts an old one, the disagreement gets written down instead of lost. When you ask a question, the answer draws on everything at once — and if it is a good answer, it gets filed too. The longer you feed it, the sharper it gets.
No. The one thing you install is Claude Code; the desktop app is fine. You never write code and never see any — you talk to it in plain sentences, in whatever language you think in. The only file you ever touch is the short rules file at the top of the folder, which you change by asking: stop making a page for every minor name, or check with me before filing anything in Spanish. Two promises are built into those rules — it never touches your originals, and every claim in the wiki has to trace back to something you gave it, or it does not get written.
Install Claude Code, open it, and paste this. That is the whole install.
Set up my second brain. Follow these steps exactly.
Rule for every step: never overwrite, replace, or delete anything
that already exists. When in doubt, stop and ask me.
1. Pick the location. If the current folder is empty or clearly meant
for this (a name like "second-brain", "brain", "notes"), build here.
Otherwise create a folder called "second-brain" in my home directory.
If that folder already exists and has files in it, stop and ask me
where to build instead.
2. Create this structure, creating only what is missing:
raw/
wiki/sources/
wiki/concepts/
wiki/entities/
wiki/index.md - heading "# Index" plus a line noting the wiki is empty
wiki/log.md - heading "# Log" plus one entry:
"## [today's date] bootstrap - second brain created"
3. If a file called CLAUDE.md already exists at the root of that folder,
leave it untouched and tell me — do not replace or merge it. Otherwise
create it with exactly this content, between the markers (do not
include the marker lines):
----- BEGIN CLAUDE.md -----
# My Second Brain - Operating Manual
This file is the schema. Any Claude session opened in this folder
follows it. It evolves: when a convention stops working, we change
it together.
## What this is
A personal wiki that Claude maintains and I feed. I supply sources
and judgement; Claude does the reading, summarising, filing and
cross-linking. I never edit wiki pages by hand - if I want to,
the schema needs updating.
## The three folders
- raw/ - I drop sources here. READ-ONLY for Claude: never modify,
rename or delete anything in it. It is the permanent original of
every source.
- wiki/ - Claude writes everything here: sources/ (one page per
ingested article, video or paper), concepts/ (named ideas that
recur across sources), entities/ (real people, companies,
products, tools). Plus index.md (table of contents - read it
first, always) and log.md (append-only history).
- CLAUDE.md - this file. We edit it together, by conversation.
## Page rules
- Filenames: kebab-case, .md, unique across the whole wiki.
- One topic per page. Cross-link with [[wikilinks]].
- Pages grow by adding sections, never by overwriting. If a new
source contradicts an old claim, record the contradiction.
- CRITICAL: never invent facts or citations. Everything in the
wiki traces back to a file in raw/ or something I said.
## How to ingest (when I say "file it")
1. Read the source completely.
2. Tell me the takeaways and what they connect to. Wait for my
reaction before writing anything.
3. Write one summary page in wiki/sources/ - how it relates to
what is already here, not a paraphrase.
4. Create or update the concept and entity pages it touches.
Link everything with [[wikilinks]].
5. Add new pages to wiki/index.md, one line each.
6. Append one line to wiki/log.md describing what happened.
## How to answer questions
Read wiki/index.md first, follow links to the relevant pages,
answer citing pages by name. If the answer is worth keeping,
file it into the wiki too - good syntheses compound.
## Evolving this schema
When a rule keeps getting in the way, propose a change to this
file and explain why. Never rewrite it unilaterally.
----- END CLAUDE.md -----
4. Do not install anything, and do not add any content to the wiki yet.
5. Finish by telling me where the brain lives and how to feed it:
save any article, transcript or note as a file inside raw/ — or just
paste the text or a link to you and ask you to save it there — then
say "file it". Remind me to open Claude Code inside this folder from
now on, so the rules are always loaded.
That makes the folders and the rules file, nothing else. Then feed it. If you do not know how to save something as a file, you do not need to — paste the text or a link into Claude Code and say save this into raw/, then file it. Give it a month before you judge it; the compounding is invisible at first.
One upgrade is worth doing early: install Obsidian, a free app, and open the folder with it. It draws the links between pages as a live graph, which is the closest thing I have to a picture of my own head. Watching that web thicken is what turns feeding the system from a chore into a habit.
If you build one, write me. I want to see what your graph looks like after a month.