You read a good article last month. You highlighted the important parts. You saved it somewhere — a bookmark, a note, a tab you promised to come back to.

You cannot find it now. You cannot remember what it said. The insight is gone, and you are starting from scratch on something it would have helped with.

This happens to everyone, all the time. Not because we are disorganised, but because brains are built for having ideas, not for storing them. Storage is a different job. It needs a second brain — a different system.

The system is a folder of plain text files, an app called Obsidian that makes them feel alive, and Claude Code pointed at the whole thing. You talk. Claude files, connects, and maintains. The folder gets smarter every week, without you doing any upkeep.

Andrej Karpathy — a founding member of OpenAI and one of the most respected voices in AI — published the idea in April 2026 as a pattern he called the AI Wiki. It went everywhere. The concept is simple: instead of asking AI to re-read your raw notes every time you have a question, let the AI build a proper library from those notes once — organised, cross-referenced, and indexed. Then every future question hits the library, not the mess.

Here is how to build it in about twenty minutes.

What Obsidian is (and why it matters that it is boring)

Obsidian is a note-taking app that stores everything as plain text files in a folder on your computer. That folder is called a vault. There is no cloud service required, no account to create, no special format. Your notes are plain text files — the same kind of file Claude Code already knows how to read and write.

This is the important part: Claude Code does not need a plugin, an extension, or any special setup to work with Obsidian. Your vault is a folder. Claude Code works with folders. That is the entire integration.

The one feature that makes Obsidian special for this workflow is wiki links. Type [[ and the name of another note, and Obsidian creates a link between them. Click the link and you jump there. If the note does not exist yet, clicking the link creates it. Every link works in both directions — if note A links to note B, note B knows about note A.

Open the graph view and you see every note as a dot, every link as a line. Clusters form. Isolated dots — orphan notes with no connections — stand out immediately. The graph is a map of your thinking, and it updates itself as you write.

The setup: one sentence and five minutes

Open Obsidian. Create a new vault — give it a name, pick a folder on your computer. Then open Claude Code, point it at that folder, and say:

“Set up this vault as a second brain. Create a raw folder for things I drop in, a wiki folder where you build structured pages, an index file you keep updated, a log file where you record everything you do, and a CLAUDE.md file with instructions for how the vault works.”

Claude creates everything. You can also say this out loud — click the mic icon and just talk. The whole workflow runs on spoken sentences.

Here is what Claude builds:

my-second-brain/
  raw/           ← drop articles, transcripts, highlights here
  wiki/          ← Claude builds structured pages here
  index.md       ← the table of contents (Claude maintains this)
  log.md         ← a dated record of everything ingested
  CLAUDE.md      ← instructions for Claude

The raw/ folder is where you dump things. Copy in an article. Paste a podcast transcript. Drop your Kindle highlights. Do not organise it — that is not your job any more.

The wiki/ folder is where Claude writes. Each page covers one idea, one concept, one person, one project. Pages link to each other with [[wiki links]]. Claude creates new pages when a concept appears that does not have one yet.

index.md is the front door. Claude keeps it updated — a clean table of contents with links to every wiki page, grouped by topic. Think of it as the card catalogue in a library.

log.md is the timeline. Every time Claude processes something new, it adds a dated entry: what was ingested, what pages were created or updated, what connections were found. You never write in this file. Claude does.

The instructions file

CLAUDE.md is the file Claude reads before doing anything. It tells Claude how your vault works. Here is a starting version:

# Second Brain

This vault is a personal knowledge base. You are the librarian.

## Structure
- `raw/` — unprocessed source material (articles, transcripts, highlights)
- `wiki/` — structured pages you maintain (one concept per page)
- `index.md` — table of contents (you maintain this)
- `log.md` — append-only record of all ingestion and changes

## Rules
- Every wiki page links to related pages using [[wiki links]]
- When you find a concept mentioned that has no page, create one
- When new information contradicts an existing page, note both versions
- Keep index.md current after every change
- Append to log.md with a date stamp after every session
- Write in plain language — this vault is for a human, not a machine

You can refine this over time. Add notes about what topics matter to you, what style you prefer, what kind of connections you care about. The file evolves as you use the system.

The daily operations

This is the part that matters. Every one of these is a single sentence — typed or spoken out loud — into Claude Code.

Ingest something new

Drop an article into raw/. Then:

“I just added an article to raw/. Read it, pull out the key ideas, write wiki pages for each one, link them to anything already in the vault, and update the index.”

Claude reads the article, identifies the core concepts, checks what already exists in wiki/, writes new pages for anything genuinely new, adds [[links]] between related pages, updates index.md, and logs the whole thing in log.md.

One article might touch ten existing pages. Claude links them. Connections you would never have noticed surface automatically — because Claude read everything in the vault before deciding what to link to what.

Ask the vault a question

“What have I saved about pricing strategy? Cite the specific pages.”

Claude scans the index, reads the relevant wiki pages, and gives you an answer grounded in your own collected knowledge — with links back to the pages it drew from. You are not searching. You are asking a librarian who has read every note you have ever saved.

Run a weekly clean-up

“Read every file in wiki/. Find orphan pages with no inbound links, concepts mentioned repeatedly but with no dedicated page, and any contradictions between pages. Give me a report.”

Your knowledge base stays healthy without you doing the filing. The orphan check alone — notes that exist but nothing links to them — catches ideas that fell through the cracks.

Morning briefing

“Read log.md, check what was added this week, and give me a three-paragraph summary of the new ideas and how they connect to what was already here.”

You open your laptop, run one command, and get a briefing that contextualises everything new against everything old. The vault is not just storing things. It is thinking about them.

Process a meeting or call

Paste a transcript into raw/. Then:

“Read the transcript. Pull out every decision, every action item with a name and deadline, and a three-bullet summary. Create or update wiki pages for any projects mentioned. Update the index.”

Every decision filed. Every action tracked. Nothing lost to a chat history you will never scroll back through.

Why this works where other systems fail

Most second brain projects die the same death. You start organised. You file things carefully for a week. Then maintenance piles up — updating the index, cross-referencing new notes with old ones, cleaning up orphans. That is extra work on top of a full workload. You skip it. The system degrades. Six months later you try to rebuild and the cycle repeats.

Claude breaks the cycle because maintenance is just a sentence. Reorganising your entire vault takes Claude two minutes and costs you ten words. The filing, cross-referencing, and bookkeeping that killed every previous system is now something you delegate in plain English.

Your job is to feed it interesting things and ask good questions. Claude handles everything else.

The honest limits

This is not magic, and pretending it is would waste your time.

Claude cannot improve disorganised thinking. If you dump vague notes with no clear ideas in them, you get a neatly filed collection of vague notes. The system organises what you give it. It cannot upgrade the quality of the raw material.

Large vaults hit a ceiling. Claude can only hold so much in its head at once. A vault with a hundred pages works beautifully. A vault with five thousand pages needs you to work by section — “read everything in wiki/pricing/” — rather than asking Claude to read the whole vault at once. The index helps here. Claude can read the index first and then pull only the pages that matter.

The setup uses the command-line version of Claude Code or the desktop app. If you have never typed a command before, the first five minutes will feel unfamiliar. After that, it is just talking.

It compounds, but slowly. A vault with ten pages is mildly useful. A vault with a hundred pages — after a few months of steady feeding — is a genuine second brain. The value curve is exponential, not instant. Stick with it past the first week.

What to put in it

Think about everything you consumed in the last year that disappeared:

  • The book you finished and forgot
  • The podcast that changed how you think about something
  • Articles saved at eleven at night that you never reopened
  • Research you did before a big decision that you cannot find now
  • Kindle highlights you marked and never looked at again

All of that belongs in your vault.

No existing material? Open Claude Code and just talk for twenty minutes. Your work, your goals, what you have been reading, what you have been thinking about. Claude will file the conversation into structured wiki pages. The vault does not need to be complete to be useful. It just needs to be real.

Steph Ango’s Obsidian skills

One more thing worth installing. Steph Ango — the CEO of Obsidian — published a set of Claude Code skills specifically for working with Obsidian vaults. They teach Claude how Obsidian formats its notes — wiki links, callouts, properties at the top of each file, the canvas layout, and even how to strip a web page down to clean text before saving it.

The skills live at kepano/obsidian-skills on GitHub. Tell Claude: “Install the Obsidian skills from kepano/obsidian-skills into this vault.” Claude downloads them and puts them in the right place. From the next session onward, Claude uses Obsidian conventions natively instead of writing generic notes.

Start here

  1. Download Obsidian from obsidian.md — it is free
  2. Create a new vault
  3. Open Claude Code and point it at your vault folder
  4. Tell Claude: “Set up this vault as a second brain with a raw folder, a wiki folder, an index, a log, and a CLAUDE.md file”
  5. Drop one article into the raw folder and tell Claude to ingest it
  6. Watch what happens

The first time Claude reads an article and creates three wiki pages with links to each other and updates the index — the first time you ask “what do I know about this?” and get an answer drawn from six months of saved reading — you will understand why a post about this setup got millions of views in days.

Your notes are not a graveyard. They are a library. They just needed a librarian.

The Karpathy gist that started this: Karpathy’s LLM Wiki. Steph Ango’s Obsidian skills for Claude Code: kepano/obsidian-skills.

Related articles: The Project That Never Forgets covers CLAUDE.md in depth — the same memory file that powers this workflow. The MCP Starter Kit shows how to connect Claude to Figma, Notion, and other tools alongside your vault. Context Rot explains why Claude gets worse as conversations get longer — and how to work within that ceiling.