Build an AI Second Brain with Claude Code
Set up an AI-powered knowledge base using Karpathy's LLM Wiki pattern with Claude Code and Obsidian.
Introduction
⚡️ 30 Second Summary
We all consume more information than we can remember, scattered across AI chats, note apps, and browser tabs that we never revisit.
In this project, you will build an AI-powered knowledge base using Claude Code and Obsidian that ingests your notes from multiple sources and compiles them into a searchable, cross-linked wiki maintained entirely by AI. Over time, every new note cross-links with what you already have, so the vault gets sharper the longer you use it.
What You'll Build
A structured Obsidian vault with a raw-source archive and an AI-compiled wiki, powered by a CLAUDE.md schema file that turns Claude Code into a disciplined wiki maintainer with custom slash commands.
By the end of this project, you'll have:
- 🗂️ A structured Obsidian vault with separate raw/ and wiki/ directories that keep your original sources untouched.
- 🤖 A CLAUDE.md schema plus a .claude/commands/ folder with four real Claude Code slash commands (/ingest, /query, /lint, /log).
- 🔗 A compiled wiki with cross-linked pages visible in Obsidian's graph view.
- ⚡ A working /ingest cycle that turns your raw notes into structured, cited wiki pages.
- 💎 Secret Mission: Put /query, /lint, and /log to work on your freshly ingested wiki.
Want a complete demo of how to do this project, from start to finish? Check out our 🎬 walkthrough with Maximus
Are there any prerequisites?
No prior experience is required. You will need a paid Claude account (Pro, Max, Team, Enterprise, or Console) so you can use Claude Code, plus Obsidian, which is free for personal use.
Not sure if this project is right for you? Check if it matches your goals
If you're up for a bit of a challenge, quiz yourself on the key concepts up ahead in this project.
This project is part of a series:
- Part 1: You are here!
Install Your Tools
To build our AI second brain, we need two tools working together: Obsidian as the visual interface for our wiki, and Claude Code as the AI agent that maintains it. Both tools will point at the same folder on your computer.
In this step, we'll install both tools so they're ready to work together in Step 2.
In this step, get ready to:
- Install Obsidian and create your vault.
- Install Claude Code and verify it's ready to use.
Install Obsidian and Create Your Vault
Obsidian is a free note-taking app that stores everything as plain Markdown files in a folder on your computer. That folder is called a "vault."
✔️ I already have Obsidian
You can skip straight to creating a new vault for this project.
- Open Obsidian.
- Click the vault switcher in the bottom-left corner, then click Create new vault (or choose Open another vault → Create).
- Name your vault second-brain.
- Choose a location you'll remember, such as your Documents folder.
- Click Create.
ⓧ I need to install Obsidian
- Go to obsidian.md/download.
- Download the installer for your operating system.
🍎 macOS
- Open the .dmg file and drag Obsidian into your Applications folder.
- Open Obsidian from your Applications.
🖼️ Windows
- Run the installer .exe file.
- Open Obsidian from the Start menu.
- When Obsidian opens, click Create new vault.
- Name your vault second-brain.
- Choose a location you'll remember, such as your Documents folder.
- Click Create
You should see an empty vault with no files. That's exactly what we want.
What is an Obsidian vault?
A vault is just a regular folder on your computer. Obsidian reads and writes Markdown files inside it. Because it's a plain folder, other tools (like Claude Code) can access the same files directly.
💡 Is Obsidian free?
Yes. Obsidian is free for personal use. Paid add-ons like Sync and Publish are optional and not needed for this project.
Install the Obsidian Web Clipper (Optional)
The Web Clipper is a browser extension that lets you save web pages directly into your vault. It's not required for this project, but it's handy if you want to clip articles later.
✔️ I already have the Web Clipper
You're all set! Make sure the extension is pinned to your browser toolbar so you can clip pages with one click later in the project.
ⓧ I need to install the Web Clipper
- Go to obsidian.md/clipper.
- Click the install button for your browser (Chrome, Safari, Firefox, Edge, Brave, or Arc). This takes you to your browser's extension store.
- Click Add Extension when your browser asks for permission.
- After installing, pin the Web Clipper extension to your browser toolbar so it's one click away. You can do this by opening your browser's extensions menu and clicking the pin icon next to Obsidian Web Clipper.
Why a Web Clipper?
A second brain is most powerful when you can capture ideas quickly. The Web Clipper lets you save any web page as a Markdown note in your vault with one click.
Install Claude Code and Connect to Your Vault
Claude Code is a command-line AI agent that can read, create, and edit files on your computer. We'll point it at the same vault folder so it can manage your wiki pages.
✔️ I already have Claude Code
You can skip the install step.
- Open your terminal and confirm your version:
claude --versionYou should see a version number. If the command is not found, switch to the I need to install Claude Code tab.
ⓧ I need to install Claude Code
- Open your terminal.
🍎 macOS/Linux
- Copy and paste the following command, then press Enter:
curl -fsSL https://claude.ai/install.sh | bash🖼️ Windows
- Open PowerShell.
- Copy and paste the following command, then press Enter:
irm https://claude.ai/install.ps1 | iexWindows prerequisite
Claude Code on Windows requires Git for Windows. If you don't have it, download it from gitforwindows.org first.
- Verify the installation by running:
claude --version✔️ I see a version number
You're all set.
ⓧ Command not found
That's okay! Let's troubleshoot:
- Close your terminal and open a new one, then try claude --version again.
- Check that the install command completed without errors.
- On Windows, make sure you restarted PowerShell after installing.
Still stuck?
Get help with your error or share your error with the NextWork community!
What Claude account do I need?
Claude Code requires a Pro, Max, Team, Enterprise, or Console account. A free Claude account will not work.
Your tools are installed. Next up, you'll point Claude Code at your vault and start building the structure of your second brain.
Build Your Vault Architecture
You have Obsidian and Claude Code installed, both pointing at the same folder. Nice work getting set up. Now it is time to give your vault some structure.
Before you can start importing notes or asking Claude to synthesize knowledge, you need a solid foundation. The directory structure you build now is the backbone of Karpathy's LLM Wiki pattern. The most important decision is the separation between raw/ and wiki/. Nearly every mistake in early reimplementations of this pattern traces back to blurring the two folders.
In this step, get ready to:
- Ask Claude Code to scaffold your vault's folder structure.
- Verify the new folders appear in Obsidian.
Navigate to Your Vault
- Open your terminal.
- Navigate to your vault folder.
cd ~/Documents/second-brain🍎 macOS
The default path is ~/Documents/second-brain. If you chose a different location in Step 1, navigate there instead.
🖼️ Windows
Your path may look different on Windows, for example C:\Users\YourName\Documents\second-brain. Adjust accordingly.
Ask Claude Code to Build the Structure
- Start a new Claude Code session in your vault folder.
claude- First time using Claude Code? Follow the prompts to sign in with your Claude account.
- Claude Code will ask for permission to access this folder. Select 1. Yes, I trust this folder.
- Fill in the MY_SOURCES variable in the prompt below with the source categories you want inside raw/. Use short, folder-safe names separated by commas (for example: claude-exports, notion-exports, granola-exports). Claude will create a subfolder inside raw/ for each name you list.
- Paste the completed prompt into Claude Code:
Create the following directory structure for my second brain vault:
second-brain/
├── raw/ # MY source documents, never modified by you
├── wiki/ # YOUR domain, you create and maintain these pages
│ ├── index.md # Master catalog of all wiki pages
│ ├── log.md # Append-only activity log
│ ├── concepts/ # Concept and topic pages
│ ├── projects/ # Project-specific pages
│ └── people/ # People dossiers
├── journal/ # Daily notes (Part 2)
└── content/ # Content pipeline (Part 2)
Then create these subfolders inside raw/, one per source I want to bring into this second brain:
[[MY_SOURCES="e.g. claude-exports, chatgpt-exports, notion-exports, granola-exports, articles, notes"]]
Add a brief README.md in every folder you create (including each raw/ subfolder) explaining what belongs there and, where relevant, where to export it from.When prompted, include the subfolders you'd like under raw/. Here are some ideas to steal from:
- AI chat: claude-exports, chatgpt-exports, perplexity, raycast-ai, cursor-chats, copilot-chats
- AI notes & meetings: granola-exports, mem, reflect, fathom, otter, fireflies, loom
- Docs & wikis: notion-exports, google-docs, apple-notes, evernote, bear, roam
- Reading: readwise, kindle-highlights, apple-books, pocket, instapaper, matter
- Media: podcast-transcripts, youtube-transcripts, voice-memos, substack
- Work: slack-exports, linear, jira, email-archives
- Research: pdfs, arxiv-papers, zotero
- Plus: articles (web clippings), notes (personal notes), or anything else you want
Claude Code will ask for your permission before running commands and creating files. You'll see prompts like "Do you want to proceed?" and "Do you want to create [file]?"
- Select Yes each time, or select Yes, allow all edits during this session to let Claude create all the folders and files without asking again.
Claude creates a brief README.md in each folder explaining its purpose. This matters because Obsidian only shows folders that contain at least one file.
What did Claude just build?
Two top-level folders with opposite ownership rules. This is the foundation of the LLM Wiki pattern. You own raw/ and drop source material into it; Claude Code only reads from it. Claude owns wiki/ and writes curated, interlinked pages based on what it finds in raw/. You read wiki/, Claude writes it.
┌────────────────────────┐ ┌────────────────────────┐ ┌────────────────────────┐
│ 1. YOUR MESSY NOTES │ │ 2. CLAUDE ORGANISES │ │ 3. YOUR SECOND BRAIN │
│ ──────────────────── │ │ ─────────────────── │ │ ─────────────────── │
│ │ │ │ │ │
│ raw/ │ │ Reads raw/, follows │ │ wiki/ │
│ • claude-chat.md │ ───► │ rules in CLAUDE.md, │ ───► │ • index.md │
│ • chatgpt-chat.md │ │ writes wiki/ │ │ • concepts/ │
│ • meeting-notes.md │ │ │ │ • people/ │
│ • notion-export.md │ │ Commands you run: │ │ • projects/ │
│ │ │ /ingest /query │ │ │
│ YOU drop notes here. │ │ /lint /log │ │ YOU read here, in │
│ │ │ │ │ Obsidian. │
└────────────────────────┘ └────────────────────────┘ └────────────────────────┘
scattered across turns mess into ask questions, spot
5 different apps linked knowledge patterns, never lose
an idea againWhy the raw/wiki split is the load-bearing rule
Out of everything in Karpathy's LLM Wiki pattern (the schema, the slash commands, the frontmatter), the one thing you cannot compromise on is the raw/ vs wiki/ separation. raw/ is your memory of record. wiki/ is Claude's interpretation of that memory. If Claude is ever allowed to edit raw/, its synthesis gets layered on top of your original notes and you lose the ability to trust either side. Keep them strictly separate and you can always re-ingest from a clean copy of your sources, roll back a bad session, or swap Claude out for a different model later without losing your ground truth.
Verify in Obsidian
- Switch to Obsidian and check the left sidebar for your new folders.
- If Graph View is not already open in a tab, open it with Cmd+G (macOS) or Ctrl+G (Windows). If it is already open, click its tab to bring it to focus.
The graph will be sparse for now, but it will grow as you add content in the next steps.
Folders not appearing in Obsidian?
Obsidian only shows folders that contain at least one file. Check that Claude Code created README.md files inside each folder. If they are missing, run the prompt again or ask our AI for help.
What makes a good wiki page?
Good pages are atomic (one idea each), densely linked with [[wiki-links]], cited back to files in raw/, synthesized instead of copy-pasted, and reachable from index.md. When sources disagree, good pages name the disagreement instead of picking a winner. You will encode all of this into your CLAUDE.md schema in Step 3, so Claude applies these rules automatically to every page it writes.
Your vault architecture is in place. Next up, you will create the CLAUDE.md file that tells Claude how to behave inside this vault. This is the schema that makes your second brain truly intelligent.
Define Your Schema and Slash Commands
Your vault has a clear folder structure, but Claude Code doesn't know any of the rules yet. Every time you start a new session, Claude would be starting from scratch with zero understanding of how your wiki works or what the page conventions are.
You'll fix that with two things:
- A CLAUDE.md file at your vault root that describes the rules of the system. Claude Code auto-reads this file every session, so it becomes your system's constitution.
- A .claude/commands/ folder with one file per workflow (ingest.md, query.md, lint.md, log.md). Each file becomes a real Claude Code slash command: type /ingest and it runs, tab-complete and all.
Together, CLAUDE.md is what your vault is, and the slash commands are how you operate on it.
In this step, get ready to:
- Ask Claude Code to generate a CLAUDE.md schema.
- Ask Claude Code to scaffold .claude/commands/ with four real slash commands.
- Test that your schema and commands work.
Ask Claude to Create the Schema
Your Claude Code session from Step 2 should still be running inside your vault root, which is exactly where it needs to be so Claude writes CLAUDE.md in the right place.
✔️ Claude Code is still running
You're good to go. Skip straight to the prompt below.
ⓧ I closed it
No worries, let's reopen it.
- Open your terminal.
- Navigate to your vault folder.
cd ~/Documents/second-brain- Start a new Claude Code session.
claudeNot sure you're in the right folder?
Run pwd in your terminal before claude. It should print the full path to your vault, for example /Users/yourname/Documents/second-brain. If it prints anything else, cd into the vault first.
- Paste the prompt below into Claude Code and press Enter.
Create a CLAUDE.md file for my second brain vault with these four sections:
1. Project Structure
Document the raw/ vs wiki/ folder separation, every subfolder,
and the purpose of index.md and log.md.
2. Page Conventions
Every wiki page must have YAML frontmatter with these fields:
title, type (one of: concept, entity, source-summary, comparison,
project, person), sources, related, created, last-updated.
Pages should use [[wiki-links]], be atomic (one idea per page),
and use a consistent heading structure.
3. Style Guide
Use clear, concise prose. Prefer bullet points. Attribute every claim
to its source. Note contradictions explicitly.
4. Domain Context
IMPORTANT: Do NOT write this section yet. First, pause and ask me
2-3 short questions, one at a time, to learn what I actually work
on day to day. Suggested questions:
- What topics or subjects come up most often in your conversations and notes?
- Who are the main people, projects, or systems you think about?
- What are you trying to remember or get smarter about over time?
Wait for my answers, then use them to write the Domain Context
section in 4-6 sentences using my own language. Do not invent
details I did not give you.Why does Claude ask questions first?
The Domain Context section needs to reflect what YOU actually work on, not a template guess. Instead of asking you to articulate your domain cold, Claude pauses and asks about your topics, decisions, and what you want to remember. Answer each in a sentence or two and Claude writes the section in your own words.
- Answer Claude's questions as they come up.
- Once you have finished answering, Claude will write CLAUDE.md to your vault root. You will see "Created CLAUDE.md" (or similar) in the terminal.
What did Claude just write?
One file at the vault root called CLAUDE.md, with four sections that together teach Claude the rules of your vault. Project Structure locks in the raw/ vs wiki/ ownership so Claude never writes into your source folders. Page Conventions sets frontmatter and linking rules. Style Guide sets the voice, including how contradictions get flagged. Domain Context names the subject matter of your work. In Karpathy's LLM Wiki pattern, this file is the spec Claude writes against. Edit it and every future page Claude writes will follow the new rules.
💡 How does Claude Code find CLAUDE.md?
When you run claude from a folder, that folder is your "current working directory." Claude Code checks it for CLAUDE.md, then the parent folder, then its parent, all the way up to your home folder. Every file it finds gets stitched into one set of instructions for the session. That is why CLAUDE.md needs to be in the vault root. It is the folder you will launch Claude from every time.
Review the Generated CLAUDE.md
Claude's first draft is not always complete. Do a fast scan for each required piece.
- In Obsidian, click CLAUDE.md in the left sidebar to open it.
Create Your Slash Commands
CLAUDE.md teaches Claude the rules of your vault. Now you need the verbs, the things you'll actually type to operate on the vault. In Claude Code, a slash command is just a Markdown file at .claude/commands/<name>.md: its contents become the prompt when you type /<name>.
- Confirm your Claude Code session is running from your vault root (the same folder that holds CLAUDE.md, raw/, and wiki/). If you're not sure, type /exit, cd into the vault, and run claude again.
- Paste this prompt into your Claude Code session:
Create a .claude/commands/ folder inside my current working directory (this vault root, next to CLAUDE.md, raw/, and wiki/). Do NOT create it anywhere else. Add four slash-command files inside that folder. Each file should start with a short frontmatter block (description, argument-hint) and then the prompt body.
1. .claude/commands/ingest.md
Description: Ingest new files from raw/ into wiki/.
Prompt body should instruct Claude to:
- For each unread file in raw/ (skip anything already summarised):
read it, write a source-summary page in wiki/, create or update
concept/project/person pages as needed, cross-link with [[wiki-links]],
update wiki/index.md, and append a timestamped entry to wiki/log.md.
- Process 5-10 sources thoroughly per run. If $ARGUMENTS is provided,
limit to that many sources.
2. .claude/commands/query.md
Description: Synthesise an answer from the wiki.
Argument-hint: the question to answer
Prompt body should instruct Claude to:
- Read wiki/index.md and the pages most relevant to $ARGUMENTS.
- Synthesise an answer grounded in the wiki, cite every claim by
wiki page name, and flag if sources disagree.
- If the synthesis reveals a new connection, propose a wiki update
but do not write it without confirmation.
3. .claude/commands/lint.md
Description: Run a health check on the wiki.
Prompt body should instruct Claude to:
- Scan wiki/ for broken [[wiki-links]], orphan pages, pages missing
required frontmatter, stale pages (30+ days untouched), and
contradictions between pages.
- Report findings as a structured list. Do not fix anything yet,
ask for permission first.
4. .claude/commands/log.md
Description: Append a timestamped note to wiki/log.md.
Argument-hint: the thought or note to capture
Prompt body should instruct Claude to:
- Append a timestamped entry containing $ARGUMENTS to wiki/log.md.
- If the note mentions a project, person, or concept that has a
wiki page, update that page too. Otherwise do not create new pages.
Create all four files now.- Approve each file creation when Claude Code asks.
Why .claude/commands/ instead of putting this in CLAUDE.md?
CLAUDE.md is context that Claude reads passively every session. Files under .claude/commands/ are real Claude Code slash commands: typing /ingest at the start of a message runs the file as a prompt, with $ARGUMENTS filled in from anything you type after the command. They tab-complete, they're shareable via Git, and they don't conflict with Claude Code's parser.
Restart Your Claude Code Session
Claude Code only scans .claude/commands/ when it starts. To make your new slash commands available, you need to restart the session.
- In your Claude Code session, type:
/exit- Then start a fresh session from the same vault folder:
claudeWhy restart?
New slash commands are not hot-reloaded. Without a restart, typing /ingest, /query, /lint, or /log will fail because the session still has the old (empty) command list.
Test That Everything Works
- Confirm your setup in a single message. Paste this into Claude Code:
List the slash commands available in this vault and tell me the rules from CLAUDE.md in one paragraph each.Claude should list /ingest, /query, /lint, and /log with their descriptions, and summarise the Project Structure, Page Conventions, Style Guide, and Domain Context from your schema.
Claude doesn't list the commands?
Check that .claude/commands/ sits at the vault root, next to CLAUDE.md, raw/, and wiki/. If the folder is there but the commands aren't picked up, restart Claude Code (/exit, then claude again) so it re-scans the commands directory. Still stuck?
Your schema and commands are meant to evolve
CLAUDE.md and your .claude/commands/ files are yours to edit freely. As you use the system, you'll discover better page conventions and new workflows worth turning into commands. This co-evolution between you and the LLM is what makes the system get smarter over time.
Your schema is locked in and your slash commands are live. Next up, you'll feed the vault real source material and put /ingest to work.
Export and Import Your Existing Data
Your vault has a solid structure, your CLAUDE.md schema is in place, and your slash commands are ready. But the vault is still empty.
The whole point of an AI second brain is to avoid starting from a blank canvas. You already have valuable knowledge scattered across Claude conversations, ChatGPT chats, Granola meeting notes, and Notion pages. In this step, you will export that knowledge and feed it into your vault so Claude Code has real material to work with.
In this step, get ready to:
- Export your existing notes and conversations from the tools you use.
- Convert JSON exports to clean markdown using Claude Code.
- Verify your imported files in Obsidian.
Export Your Conversations
You do not need all of these sources. Even 5-10 items from a single tool is enough to get started. Pick the tool you use most and come back for the others later.
Claude
- Open claude.ai in your browser.
- Click your initials in the lower-left corner.
- Select Settings.
- Navigate to Privacy.
- Click Export data.
- Leave the default settings (export All conversations) and click Export.
You will receive a download link via email within a few minutes. The link expires in 24 hours.
- Open the email and download the ZIP file.
- Unzip the file.
What is in the export?
The Claude export contains JSON files with your conversation history. Each file holds the messages from a conversation, including both your prompts and Claude's responses.
- Drag the unzipped export folder from your Downloads into the Claude Code terminal. This pastes the folder path.
- After the path, type move all these files into raw/claude-exports/ and press Enter.
- Claude Code will ask for permission to access the folder. Select Yes to proceed.
Your Claude export files are now in raw/claude-exports/. You won't see them in Obsidian's sidebar yet because Obsidian hides JSON files by default. You'll convert them to readable Markdown in the next section.
Export option not showing up?
The export feature is only available from the Claude web app. It is not available on mobile. Make sure you are signed in at claude.ai in a desktop browser. Still stuck?
ChatGPT
- Open chat.openai.com in your browser.
- Click your profile icon.
- Select Settings.
- Navigate to Data Controls.
- Click Export Data.
- Confirm via email.
You will receive a ZIP file via email. The link expires in 24 hours.
- Download the ZIP file from the email.
- Unzip the file.
What is in the export?
The ChatGPT export contains a conversations.json file with your full chat history. Each conversation includes metadata like the title and timestamps alongside the messages.
- Drag the unzipped export folder from your Downloads into the Claude Code terminal.
- After the path, type move all these files into raw/chatgpt-exports/ and press Enter.
Your ChatGPT export files are now in raw/chatgpt-exports/. You won't see them in Obsidian's sidebar yet because Obsidian hides JSON files by default. You'll convert them to readable Markdown in the next section.
Email not arriving?
Check your spam folder. The export email can take up to 10 minutes to arrive. Still waiting?
Granola
Granola doesn't have a built-in "export everything" button, so you have two options. Pick whichever fits.
⚡ Export every meeting at once
Granola caches your signed-in credentials locally, so Claude Code can script a bulk export of every meeting straight into your vault. No clicking through the app.
- In your Claude Code session (running from your vault root), paste:
I want to bulk export every meeting note from Granola into raw/granola-exports/ as individual Markdown files.
Please:
1. Check that Granola is signed in on this machine by looking for ~/Library/Application Support/Granola/supabase.json (macOS) or the equivalent on Windows/Linux. Tell me what you find.
2. Use the credentials from that file to call Granola's get-documents API and fetch every meeting I have access to.
3. For each meeting, write a Markdown file into raw/granola-exports/ containing the AI-generated notes, attendees, date, and transcript if available. Name the file using the meeting date and title.
4. Add YAML frontmatter with: source: granola, date, attendees, and topic.
5. Before running anything, summarise the plan and ask me to confirm. After running, tell me how many meetings were exported and flag any that failed.- Claude Code will ask for permission to read the credentials file and make network requests. Select Yes when it does.
Why this works
Granola stores an authenticated session token on disk when you sign in. Claude Code reads that token, calls Granola's API the same way the desktop app does, and writes every meeting as Markdown. Your credentials never leave your machine.
🙋♀️ Claude Code says it can't find the credentials file?
Make sure you're signed into the Granola desktop app at least once. Ask Claude Code to troubleshoot if the file really isn't there.
🎯 Export a few meetings I care about
If you only want the meetings that actually mattered, export them one at a time from the app.
- Open the Granola desktop app.
- Find a meeting note you want to export in the left sidebar.
- Right-click the note (or click the ... menu at the top of the note).
- Select Export as Markdown.
- Choose where to save the file.
Repeat this for each meeting you want in your second brain. Start with the last 10-20 meetings that actually mattered.
- Drag the folder with your exported files into the Claude Code terminal.
- After the path, type move all these files into raw/granola-exports/ and press Enter.
Can't find the export option?
Granola's menus move around between versions. Ask Claude Code for help.
What is in the export?
Either path gives you Granola's AI-generated meeting notes as Markdown files, including the summary, decisions, and action items. They are already in a readable format, so no JSON conversion is needed later.
Notion
- Open Notion in your browser.
- Click Settings & Members in the left sidebar.
- Select Settings.
- Scroll down and click Export all workspace content.
- Choose Markdown & CSV as the export format.
- Click Export.
You will receive a download link via email.
- Download the ZIP file from the email.
- Unzip the file.
What format does Notion export?
Notion exports pages as Markdown files. Unlike Claude and ChatGPT exports, these are already in a readable format, so no conversion is needed later.
- Drag the unzipped export folder from your Downloads into the Claude Code terminal.
- After the path, type move all these files into raw/notion-exports/ and press Enter.
Pro tip
Exporting individual pages
You can also export specific pages instead of your entire workspace. Open a page, click the ... menu in the top right, then select Export and choose Markdown & CSV.
🤔 Not sure how to get your data?
Don't see your tool in the tabs above? No problem. Almost every knowledge app has some way to export your data, but the exact steps change every few months. Instead of hunting through menus, let Claude Code do the detective work for you.
- Create a new folder inside raw/ for your tool. For example, if you use Apple Notes, create raw/apple-notes-exports/.
- Open Claude Code in your terminal from your vault root.
- Paste this prompt and replace enter your tool name with the tool you want to export from:
I want to pull my data out of [[TOOL_NAME="enter your tool name"]] and save it into raw/[[TOOL_NAME="enter your tool name"]]-exports/ inside my Obsidian vault so I can feed it into my AI second brain.
Please help me by:
1. Looking up the current, official export instructions for [[TOOL_NAME="enter your tool name"]] and summarising the exact steps I need to click through.
2. Telling me what format the export will come in (JSON, HTML, CSV, Markdown, PDF, etc.) so I know what I'm working with.
3. If [[TOOL_NAME="enter your tool name"]] has an API, CLI, or scriptable export, offer to write a small script that downloads my data directly into raw/[[TOOL_NAME="enter your tool name"]]-exports/ for me.
4. Once I've got the files, telling me whether I need a conversion step to turn them into clean Markdown for the ingest process.
Start by telling me what you found about [[TOOL_NAME="enter your tool name"]]'s export options before writing any code.- Follow the steps Claude Code gives you and drop the exported files into the folder it suggests.
Tools that work well as sources
Apple Notes, Google Docs, Obsidian (another vault), Bear, Roam Research, Logseq, Readwise highlights, Todoist, Linear, and even voice memo transcripts all make great raw material for a second brain. If a tool stores your thinking, it's fair game.
🙋♀️ Claude Code can't find export instructions?
Some smaller tools do not document their export flow clearly. Ask Claude Code to inspect the app's local storage as a fallback.
Add as many sources as you'd like until you're done. You only need one source to get started, but the more you import, the richer your second brain will be. Scroll back up and click through the other tabs to export from each tool you use.
Convert Exports to Clean Markdown
Some exports arrive in JSON (Claude, ChatGPT, Perplexity, Raycast AI, Cursor, Copilot, Readwise, and many meeting tools like Fathom, Otter, or Fireflies), which is not human-readable. Others arrive already in Markdown (Notion, Granola, Apple Notes) and don't need conversion.
The prompt below lets Claude Code scan every raw/ subfolder, find whatever JSON you imported, and convert only those files. If none of your sources are JSON, Claude will tell you there's nothing to convert and you can skip ahead.
Closed your Claude Code session?
If you closed Claude Code after Step 3, reopen it with claude from your vault root before continuing.
- Paste this prompt into your Claude Code session:
Scan every subfolder inside raw/ for JSON files. For each JSON file you find:
1. Parse the conversations, messages, highlights, or records inside
2. Convert each item to clean markdown
3. Name each file based on the conversation topic, title, or first message
4. Save the converted files back in the same raw/ subfolder alongside the original JSON
5. Add YAML frontmatter to each file with: source (the subfolder name), date, and topic
If you find a JSON structure you don't recognise, tell me what you see before writing the script so we can decide the best conversion together. If there are no JSON files in raw/, tell me which subfolders you checked and stop - no conversion needed.
Then run the script.- Press Enter and let Claude Code generate and execute the script.
What does this script do?
The script reads raw JSON conversation files, extracts the messages, and writes each conversation as a clean Markdown file with YAML frontmatter at the top. The frontmatter tags each file with its source (Claude or ChatGPT), the date, and the topic, which makes it easier for the ingest process to organize later.
✔️ It worked
You should see new .md files appearing in your raw/ subfolders. Each one is a converted conversation.
ⓧ I see an error
That's okay! Let's troubleshoot:
- Tell Claude Code exactly what went wrong. Paste the error message back into the chat:
The script failed - here's the error: [paste your error here]. Can you fix it and try again?- The JSON format for these exports can change over time. If the script cannot parse your files, ask Claude Code to inspect the raw file first:
Can you look at the first 50 lines of raw/claude-exports/[filename].json and figure out the structure, then rewrite the conversion script?Still stuck?
Get help with your error or share your error with the NextWork community!
Verify Your Files in Obsidian
- Switch to Obsidian and open the file explorer in the left sidebar.
- Expand the raw/ folder and its subfolders.
- Confirm that you can see .md files in the export subfolders.
- Open a few of the converted files.
- Check that they have readable content and YAML frontmatter at the top. In Obsidian, frontmatter is auto-rendered as a Properties panel right under the file title, with one row per field (e.g. source, date, topic). If you see that Properties panel, the YAML is there.
Want to see the raw YAML?
Obsidian hides the --- block and shows it as the Properties pane by default. To see the underlying YAML, press Cmd+E (macOS) or Ctrl+E (Windows) to toggle Source mode. You'll see something like this at the very top of the file:
---
source: claude-exports
date: 2026-03-26
topic: Building muscle while cutting fat and improving mobility
---Do not worry about perfect formatting at this stage. The ingest process in the next step handles synthesis and cross-linking.
Large exports? Start small
Claude Code has context window limits. If you exported hundreds of conversations, start with 10-20 of your most relevant ones. You can always add more later by running the conversion script again on the remaining files.
Your vault now has real content to work with. Next up, you will run the /ingest command and watch Claude Code compile your raw notes into a cross-linked wiki.
Run Your First Ingest Cycle
Your raw/ folder is loaded with real content, your CLAUDE.md defines the rules, and your .claude/commands/ingest.md is waiting to be invoked. Now it is time to run the core operation that makes the entire system work: the ingest.
/ingest is the heart of Karpathy's pattern. Instead of searching through scattered raw notes every time you need something, Claude pre-compiles your knowledge into structured wiki pages with cross-links. Think of it like a compiler turning source code into an executable. Your messy notes go in, and a navigable knowledge base comes out.
In this step, get ready to:
- Run /ingest to process your raw sources.
- Explore the wiki pages Claude Code compiled.
- View your knowledge graph in Obsidian's Graph View.
Start the Ingest
Closed your Claude Code session?
If you closed Claude Code after Step 4, reopen it with claude from your vault root before continuing.
🚨 Heads up: /ingest eats tokens
A full ingest reads every file in raw/ and writes new pages into wiki/, which burns through a lot of tokens and can take 10+ minutes on a large batch. On a Claude Pro plan, big ingests can push you into the 5-hour usage limit. On the API, a heavy run can cost a few dollars.
The fix is simple: start with a tiny demo ingest, then run the full ingest in small batches. /log and /query are cheap by comparison, so lean on those day-to-day and reserve /ingest for when you have new source material.
Start with a tiny demo ingest so you can see the workflow in action before committing to a long run.
- In Claude Code, type:
/ingest 2The 2 is passed in as $ARGUMENTS to your ingest command, telling Claude to only process 2 sources. A minute or two later you'll see source-summary pages, concept pages, cross-links, and updated index.md / log.md appear in Obsidian. That's a working slice of the whole system on a small sample.
What /ingest is actually doing
For each file in raw/, Claude:
• Writes a summary of it in wiki/
• Pulls out the concepts, people, and projects inside into their own pages
• Links everything together with [[wiki-links]]
• Updates index.md (your table of contents) and log.md (your activity log)
The full recipe lives in .claude/commands/ingest.md. Edit that file any time you want to change how /ingest behaves.
BEFORE AFTER
──────────────────────── ────────────────────────
raw/ wiki/
└── claude-exports/ ├── index.md (catalog)
├── chat-on-pricing.md ──┐ ├── log.md (activity)
├── chat-on-hiring.md ──┤ ├── concepts/
└── chat-on-launch.md ──┤ ──/ingest──► │ ├── pricing.md
└── granola-notes/ │ │ └── hiring.md
├── standup-tues.md ──┤ ├── people/
└── 1-1-with-alex.md ──┘ │ └── alex.md
└── projects/
└── launch.md
one file per conversation, one page per idea,
one file per meeting all cross-linked with [[wiki-links]]Once the demo looks right, run /ingest without arguments to process the rest of your sources in batches.
- When you're ready to compile the rest, type:
/ingestThis runs the command with no argument, so it uses the default "5-10 sources thoroughly" rule from your ingest.md file. Re-run it as many times as you need to work through raw/.
Pro tip
Watch your graph grow in real time
Open Obsidian's Graph View by pressing Cmd+G (macOS) or Ctrl+G (Windows) while the ingest runs. You will see new nodes and connections appear as Claude creates each wiki page.
Claude will work through your raw sources one by one. For each file, it reads the content, creates structured wiki pages, and cross-links everything using [[wiki-links]]. This process may take a few minutes depending on how many sources you have.
Explore the Results
- Open wiki/index.md in Obsidian. This is the master catalog of every page Claude created.
- Open wiki/log.md to see a timestamped record of everything Claude did during the ingest.
- Browse through the wiki subfolders to explore what Claude compiled.
The wiki/concepts/ folder holds the key ideas and frameworks Claude pulled from your sources. The wiki/people/ folder contains dossiers for anyone mentioned in your notes. The wiki/projects/ folder holds the projects and initiatives that came up across your sources.
- Click any [[wiki-link]] inside a page to navigate between connected pages.
What are wiki-links?
Wiki-links use the [[page-name]] syntax to create connections between pages. When you click a wiki-link in Obsidian, it jumps straight to the linked page. These links are what turn a folder of isolated notes into a connected knowledge graph.
Explore Graph View
- Switch to your Graph View tab in Obsidian. If it is not open, press Cmd+G (macOS) or Ctrl+G (Windows).
You should now see a web of interconnected nodes. Each node represents a wiki page and each line represents a [[wiki-link]] between pages. Clusters of tightly connected nodes indicate related topics.
What am I looking at?
The graph is the visual payoff of the ingest cycle. As Andrej Karpathy puts it, "Obsidian is the IDE, the LLM is the programmer, the wiki is the codebase." The dense clusters you see are topics where your notes had the most overlap. Sparse nodes may need more source material.
🙋♀️ Only seeing a few pages or links?
Check that your raw/ subfolders actually contain markdown files. If you imported a large number of sources, try running the ingest in batches by subfolder. Even 5-10 conversations is enough to see meaningful connections. How can I check my raw folder contents?
Your second brain has real compiled knowledge. That's a complete system you can start using today. If you want to push further and put your other three slash commands to work, the Secret Mission below walks through /query, /lint, and /log on your freshly ingested wiki.
Secret mission
You've got /ingest running. In this secret mission, you will put the other three slash commands through their paces: /query to pull synthesised answers out of your wiki, /lint to health-check it, and /log to capture ideas on the fly.
In this secret mission, get ready to:
- Use /query to synthesise an answer from your wiki.
- Use /lint to health-check the wiki and fix any issues.
- Use /log to capture a thought that isn't worth a full source.
THE FOUR VERBS OF A SECOND BRAIN
────────────────────────────────
/ingest raw/ ──► wiki/ compile sources into pages
/query wiki/ ──► answer ask questions, get cited answers
/lint wiki/ ──► report find broken links & orphan pages
/log note ──► log.md capture an idea in one line
These four are the only things you ever do inside your second brain.Put Your Four Verbs to Work
Clean Up Your Resources
Clean Up Your Resources
Decide whether to keep your resources running, pause them to come back later, or delete them entirely. This project runs entirely locally, so there are no ongoing costs.
Resources you used:
- Obsidian vault directory (your second brain).
- CLAUDE.md configuration file inside the vault.
- raw/ folder with the notes and exports you imported from your everyday tools.
- wiki/ folder with ingested notes.
- log.md session log.
- Claude Code session.
✔️ Keep everything running
No action needed. Choose this if you're still actively building or want to keep testing right away.
Your vault and all its files are stored locally with no ongoing costs. The CLAUDE.md file, raw/ exports, and wiki/ notes will stay exactly where they are.
Back up your CLAUDE.md
CLAUDE.md is the brain of your second brain. Copy it to cloud storage (iCloud, Google Drive, Dropbox) or a private GitHub repo so you never lose your custom instructions.
✋ Pause - I'll come back to this later
Shut down running processes to free up memory, but keep all your files and data so you can pick up where you left off.
- Type /exit in your Claude Code session to close it cleanly. This ensures any in-progress file writes complete and log.md is flushed.
Your vault files, CLAUDE.md, and all exports in raw/ will remain on your machine. To pick up later, open Obsidian, navigate to your vault, and start a new Claude Code session.
Back up your CLAUDE.md
Before stepping away, copy CLAUDE.md to cloud storage or a private Git repository. Your vault is plain Markdown, so it works perfectly with Git. Run git init inside your vault folder and commit regularly for version history, backup, and rollback capability.
ⓧ Delete - I don't want to use this again
Remove all project resources and start fresh if you ever want to rebuild.
- Type /exit in your Claude Code session if it is still running.
- Delete the entire vault directory:
rm -rf ~/second-brain- If you installed Obsidian only for this project and no longer need it, uninstall it through your system's application manager.
The raw/ folder is immutable
If you only want to redo your wiki pages, you do not need to delete everything. Update CLAUDE.md with new instructions and re-run the /ingest command. Never delete files from raw/ directly.
That's a Wrap!
That's a Wrap!
Nice work! 🚀 You've just built an AI-maintained knowledge base that turns scattered notes into a structured, cross-linked wiki using Claude Code and Obsidian. This is Part 1 of a 3-part AI Second Brain series.
You've learned how to:
- 🏗️ Set up a structured Obsidian vault using Karpathy's raw/ and wiki/ architecture.
- 📥 Import your existing knowledge from your everyday tools (Claude, ChatGPT, Granola, Notion, or whatever you use) into raw/ as Markdown files.
- 📝 Write a CLAUDE.md schema and author four real Claude Code slash commands (/ingest, /query, /lint, /log) under .claude/commands/.
- 🔄 Compile a wiki with cross-linked concept pages, people dossiers, and project pages using /ingest.
- 🌐 Explore your knowledge as an interconnected web using Obsidian's Graph View.
- 💎 Put /query, /lint, and /log to work in the Secret Mission to turn your vault into a system you use daily.
Ready to quiz yourself? 💪
p.s. Does it say "Still tasks to complete!" at the bottom of the screen?
This means you still have screenshots left to upload, or questions left to answer!
- Press Ctrl+F (Windows) or Command+F (Mac) on your keyboard.
- Search for the text Return to later.
- Jump straight to your incomplete tasks!
- 🙋♀️ Still stuck? Ask the community!