How to build a second brain with Claude in ten minutes.

The fix is a folder.

Most people use Claude like a vending machine. Paste context, get answer, close the tab, start over tomorrow. Nothing compounds.

The fix is a second brain. One folder Claude reads from and writes back into, so every source you feed it makes the next answer smarter.

This guide gives you the working system in two flavors. Use claude.ai if you live in the browser. Use Claude Code if you live in the terminal. Pick one and ship it today.

Your folder becomes your app.

The pattern comes from Andrej Karpathy. The shape is simple. You drop sources into one place. Claude reads them once, extracts what matters, and writes it into a second place as plain markdown notes that link to each other.

From then on, when you ask a question, Claude reads the notes, not the raw sources. Knowledge is compiled once and kept current. No vector database. No embeddings. No pipeline. The folder is the app.

You only need three operations to make this work.

Three layers, every time.

The map. A single instructions file Claude reads first, every time. It tells Claude where things live and which file to open for which kind of question. Think of it as the floor plan posted at the entrance.

The rooms. Two folders. raw/ holds your immutable sources. Claude reads them but never edits them. wiki/ is Claude's workspace. This is where the synthesis lives.

The pages. Plain markdown files inside wiki/ . Each page covers one entity, concept, or analysis. Pages link to each other with double brackets. An index page lists them all. A log page keeps a running record of every operation.

The browser version, in Projects.

Open claude.ai. Create a new Project. Name it whatever you want. Your second brain.

Paste this into the Project's custom instructions.

You are the librarian of a second brain. The Project Knowledge holds the wiki. Treat it as the source of truth.
Your three jobs:
1. INGEST. When I add a new source (PDF, transcript, link, pasted text), read it once and produce or update wiki pages. Each page covers one entity or one concept. Use [[double brackets]] to link related pages. Keep an index page that lists every page. Keep a log page with one dated line per operation.
2. QUERY. When I ask a question, read the wiki first. Answer from the wiki. If the wiki is missing the answer, say so before guessing.
3. HEAL. When I say "run a self heal," audit the wiki. Find pages that contradict each other. Find concepts mentioned in multiple pages but never given their own page. Find broken links. Report them. Then, only on my approval, fix them.
Rules.
- One idea per page.
- Short pages beat long pages.
- Always update the index when you add or rename a page.
- Never invent sources. If something is not in the wiki or the raw upload, say so.
- When you cite a fact, name the source page or upload it came from.

Paste into your Claude Project's custom instructions, exactly once.

Now upload your first source as a Project Knowledge file. A PDF works. A meeting transcript works. A long article you want to remember works.

Tell Claude this.

Ingest the source I just uploaded. Build the first wiki pages. Create the index and the log. Show me the page list when you're done.

Paste into the Project chat after your first upload.

That's the brain. Every new chat in this Project starts smarter than the last because the wiki travels with you.

The terminal version, with two skills.

This is the path the repo was built for. It runs in Claude Code on your machine and gives you git history, scheduled self-heal runs, and a folder you can open in any markdown editor.

Install the two skills.

git clone https://github.com/NulightJens/ai-second-brain-skills.git ~/ai-second-brain-skills
ln -s ~/ai-second-brain-skills/llm-wiki-setup  ~/.claude/skills/llm-wiki-setup
ln -s ~/ai-second-brain-skills/wiki-self-heal  ~/.claude/skills/wiki-self-heal

Run these three lines in your terminal. Restart Claude Code after the first install.

Set up your vault.

Set up an LLM wiki at ~/second-brain

Type this into Claude Code from any directory.

Drop your first source into ~/second-brain/raw/ . Then from inside the vault.

Ingest the new source I just added to raw/

Run from inside the vault folder.

Once a week, run the heal step.

Run a dry-run audit of this vault. Don't change anything yet.

First run is always a dry run. Read the report before you let it touch the files.

Three commands, forever.

You only need to remember three things. Add a source. Ask a question. Heal once a week.

The compounding shows up around the fifth source. By then the wiki is dense enough that Claude answers cross-source questions. "Where do my sources agree and disagree about X." "What's the most common counter-argument to Y." "Write me a comparison page for A versus B and file it back into the wiki."

Good answers get filed back as new pages. Exploration compounds along with sources.

Five mistakes to skip.

1. Skipping the instructions file. Without the map, Claude wanders. Paste it before you upload your first source, not after.
2. Treating raw sources as editable. Claude should never change the raw folder. If a source is wrong, add a correction page in the wiki and link to it.
3. Pasting context into the chat instead of feeding it to the brain. Every time you do this, the knowledge dies when the chat closes.
4. Auto-running self heal before you've run it manually once. The first audit always finds more than you expect. Read it first.
5. Building a giant wiki for a tiny use case. This pattern shines for personal research, ongoing projects, and small teams. If you have millions of documents, use a real vector database.