We make services people love by the power of Gen AI.
Facebook-f Twitter Youtube Linkedin-in
Contact

Building a Second Brain “LLM Wiki” with Claude Code and Obsidian

Part 1 — Concept: What Is an “LLM Wiki”?

The Core Idea

A traditional personal wiki (like Notion, Roam, or a plain Obsidian vault) stores your knowledge in files. But it has a fundamental bottleneck: you still have to search, read, and connect the dots manually. The bigger the wiki grows, the harder it is to extract value from it.

Andrej Karpathy’s “LLM Wiki” concept flips this model. Instead of you querying your own notes, an LLM is given continuous access to your entire knowledge base and acts as a conversational interface over it. Your notes become the LLM’s long-term memory. You stop searching and start asking.

The key insight:

A wiki is only useful if it is queryable. An LLM can read all your notes at once and synthesize answers — this turns a passive archive into an active intelligence layer.

Architecture in Plain Terms

┌─────────────────────────────────────┐
│         Obsidian Vault              │
│  (your notes, written by you)       │
│                                     │
│   CLAUDE.md  ← "system prompt"      │
│   Projects/  ← active work          │
│   Areas/     ← life domains         │
│   Resources/ ← knowledge base       │
│   Daily/     ← daily logs           │
│   Inbox/     ← quick captures       │
│   Archive/   ← completed items      │
└────────────────┬────────────────────┘
                 │ file system access
                 ▼
┌─────────────────────────────────────┐
│         Claude Code (CLI)           │
│                                     │
│  - Reads every .md file             │
│  - Understands WikiLinks            │
│  - Follows CLAUDE.md rules          │
│  - Creates / edits / links notes    │
│  - Answers questions from context   │
└─────────────────────────────────────┘

Role of Each Tool

Tool Role Why It Matters
Obsidian Knowledge container & visualization layer Stores notes in plain Markdown files on disk. No lock-in. Human-readable. Supports [[WikiLinks]] that Claude can parse. Graph view lets you see knowledge connections visually.
Claude Code AI agent with file system access Runs as a CLIinside your vault folder. Unlike a chat UI, it can read, create, edit, and link files directly — no copy-paste loop required.
CLAUDE.md Persistent system prompt / “brain stem” A special Markdown file read by Claude Code at startup. It defines your AI’s behavior, note templates, folder conventions, and language rules. It is the glue that makes the vault “AI-native”.

The combination is powerful because of one technical fact: Claude Code’s context window can hold your entire vault. For a typical personal knowledge base (hundreds of short notes), the AI has full awareness of everything you’ve ever written — simultaneously.

Part 2 — Setup: Step-by-Step with Execution Evidence

Environment

  • OS: macOS
  • Tools required: Node.js ≥ 18, Claude Code CLI, Obsidian (desktop app)

Step 1 — Install Claude Code

npm install -g @anthropic-ai/claude-code

Claude Code is Anthropic’s official agentic CLI. After installation, authenticate with your Anthropic API key:

claude auth login

Step 2 — Create the Vault Directory with P.A.R.A Structure

The P.A.R.A method (Projects, Areas, Resources, Archive) by Tiago Forte is the organizational backbone. Daily notes and an Inbox folder are added for capture workflow.

mkdir -p ~/llm-wiki-demo/MyBrain/{Projects,Areas,Resources,Archive,Inbox,Daily}

Execution result — vault directory created:

$ ls -la ~/llm-wiki-demo/MyBrain

total 8
drwxr-xr-x  9 tinhnguyen  staff   288 Jun 10 13:31 .
drwxr-xr-x  3 tinhnguyen  staff    96 Jun 10 13:30 ..
drwxr-xr-x  2 tinhnguyen  staff    64 Jun 10 13:30 Archive
drwxr-xr-x  3 tinhnguyen  staff    96 Jun 10 13:36 Areas
-rw-r--r--  1 tinhnguyen  staff  1279 Jun 10 13:31 CLAUDE.md
drwxr-xr-x  3 tinhnguyen  staff    96 Jun 10 13:35 Daily
drwxr-xr-x  3 tinhnguyen  staff    96 Jun 10 13:31 Inbox
drwxr-xr-x  3 tinhnguyen  staff    96 Jun 10 13:36 Projects
drwxr-xr-x  3 tinhnguyen  staff    96 Jun 10 13:35 Resources

Step 3 — Write CLAUDE.md (The AI Brain Stem)

CLAUDE.md is the most critical file. It is automatically read by Claude Code when launched inside the vault. It defines:

  • The AI’s role and language behavior
  • The folder structure so Claude always knows where to save things
  • Core rules for note hygiene (frontmatter, WikiLinks, never-delete policy)
  • A note template to enforce consistency
  • Common commands as shorthand aliases

File created: ~/llm-wiki-demo/MyBrain/CLAUDE.md

# CLAUDE.md — LLM Wiki Brain Instructions

## Role
You are my personal knowledge assistant. This Obsidian vault is my Second Brain.
Always respond in the same language the user writes in.

## Vault Structure (P.A.R.A)
- **Inbox/**     → Unprocessed notes, raw captures
- **Projects/**  → Active projects with deadlines
- **Areas/**     → Ongoing responsibilities (health, finance, career)
- **Resources/** → Reference material, topic libraries
- **Archive/**   → Completed or inactive items
- **Daily/**     → Daily notes (YYYY-MM-DD.md)

## Core Rules
1. When creating a note, always add frontmatter: `tags`, `created`, `status`
2. Link related notes using [[WikiLinks]]
3. Never delete notes — move to Archive/ instead
4. Summarize long content into bullet points
5. When asked "what do I know about X", search all files and synthesize

## Note Template
---
tags: [topic, subtopic]
created: {{date}}
status: draft | active | archived
---
# Title
## Summary
> One-line TL;DR
## Details
## Related
- [[linked-note]]

## Commands I Use Often
- "capture this"              → save to Inbox/
- "summarize vault on [topic]"→ search + synthesize
- "create project note for X" → use template in Projects/
- "what did I learn about X?" → semantic search across vault

Step 4 — Seed the Vault with Initial Notes

Using Claude Code from inside the vault, initial notes were created to populate each P.A.R.A category. This demonstrates the AI’s ability to write structured, linked notes on command.

Command run:

cd ~/llm-wiki-demo/MyBrain
claude

Prompt given to Claude Code:

Create a Resources note about the LLM Wiki concept, a Project note for 
the setup process, a Daily note for today, a raw Inbox capture, 
and an Areas note for Career. Link them together with WikiLinks.

Execution result — all files generated:

$ find ~/llm-wiki-demo/MyBrain -type f | sort

./Areas/Career.md
./CLAUDE.md
./Daily/2025-01-15.md
./Inbox/2025-01-15-capture.md
./Projects/LLM-Wiki-Setup.md
./Resources/LLM-Wiki-Overview.md

Total: 6 Markdown files across 5 folders

Every file was created with correct frontmatter, proper WikiLink cross-references, and filed in the right P.A.R.A folder — all from a single natural language prompt.

Step 5 — Open Vault in Obsidian

  1. Launch Obsidian → Open folder as vault
  2. Navigate to ~/llm-wiki-demo/MyBrain
  3. Obsidian auto-discovers all .md files

The Graph View (Cmd+G) immediately renders the link network between notes. Because Claude Code wrote WikiLinks into every file, the knowledge graph is populated automatically — no manual linking needed.

Sample note — Resources/LLM-Wiki-Overview.md (created by Claude Code):

---
tags: [ai, llm, second-brain, knowledge-management]
created: 2025-01-15
status: active
---
# LLM Wiki — Overview

## Summary
> An LLM-powered personal wiki where your AI assistant has full context 
> of everything you know.

## Core Concept (Karpathy's Vision)
- A wiki is only useful if it's **queryable**
- Traditional wikis require YOU to search and connect dots
- LLM Wiki: the AI reads all notes and **synthesizes answers** for you
- CLAUDE.md acts as the "system prompt" for your entire brain

## Related
- [[Daily/2025-01-15]]
- [[Inbox/2025-01-15-capture]]

Step 6 — Test an AI Query Across the Vault

Prompt:

> "What do I know about LLM Wiki so far? Summarize everything."

Claude Code reads all 6 files simultaneously and returns a synthesized answer — without you opening a single file. This is the core value: the AI does the retrieval and synthesis for you.

Vault State Summary

Category Files Status
CLAUDE.md 1 ✅ AI instructions active
Projects/ 1 ✅ LLM-Wiki-Setup tracked
Areas/ 1 ✅ Career skills mapped
Resources/ 1 ✅ Concept documented
Daily/ 1 ✅ Today’s log captured
Inbox/ 1 ✅ Raw capture saved
Archive/ 0 (empty — nothing archived yet)

Part 3 — Personal Application: How I Would Use This Daily

My Context

As a developer and technical writer, I constantly juggle three types of cognitive overhead:

  1. Meeting/discussion notes that I never re-read but occasionally need
  2. Learning fragments from articles, docs, and videos that I forget within a week
  3. Decision logs for architectural choices that become tribal knowledge

The LLM Wiki system directly addresses all three.

Use Case 1: The “What Did We Decide?” Problem

Pain point: A colleague asks why we chose library X over Y six months ago. I have no idea. The context lives in a Slack thread that no longer exists.

LLM Wiki solution:

  • After every design discussion, I type one prompt: "capture: we chose Zustand over Redux because of bundle size and simpler mental model for our team size"
  • Claude Code saves this to Inbox/ with today’s date and tags
  • Six months later: "why did we choose Zustand?" → Claude synthesizes the decision log from the vault instantly

Use Case 2: Learning Retention

Pain point: I read 10 articles about RAG architecture. A month later I remember almost nothing useful.

LLM Wiki solution:

  • After reading, I paste key excerpts to Claude Code: "capture my notes on RAG chunking strategies"
  • Claude formats it into Resources/RAG-Chunking.md with proper tags and links it to related notes ([[Resources/LLM-Wiki-Overview]])
  • Later: "what do I know about RAG?" → instant synthesis of everything I’ve captured, with source links

Use Case 3: Weekly Review Automation

Pain point: My Friday review takes 30–40 minutes of manually re-reading notes to remember what happened.

LLM Wiki solution:

> "Summarize this week's Daily notes and list open tasks"

Claude reads all daily logs from the past 7 days and returns a structured summary in under 10 seconds. Review time drops from 40 minutes to 5.

Use Case 4: Onboarding New Team Members

Pain point: Onboarding docs are always out of date because nobody wants to update them.

LLM Wiki solution: If the team vault is shared (git-synced), an onboardee can ask:

> "What should I know about our backend architecture?"

Claude synthesizes from all active notes across Projects, Areas, and Resources — always reflecting the current state of the vault.

Expected Productivity Impact

Task Before After Saving
Finding a past decision 10–20 min (Slack search) < 30 sec ~95%
Weekly review 30–40 min 5 min ~85%
Article → usable note Never happened 2 min ∞
Answering “what do we know about X” Unknown < 1 min Significant

One Honest Limitation

The system requires discipline in capturing. Claude Code is only as smart as the notes you feed it. If you don’t capture, there’s nothing to synthesize. The habit of capturing (even rough Inbox notes) is the real bottleneck — not the technology.

The practical fix: lower the bar for capture. A one-sentence Inbox note is infinitely better than a perfect note you never write.

Conclusion

The LLM Wiki is not a new app — it is a new way of relating to your own notes. Obsidian provides the frictionless Markdown storage and visual graph. Claude Code provides the agentic AI layer with direct file system access. CLAUDE.md is the glue that makes it feel like a persistent, personalized second brain rather than a generic chat session.

The setup takes under 15 minutes. The compounding value grows with every note you add.

The best productivity system is the one you actually use — and asking a question is always easier than searching a folder.

Related

  • [[Resources/LLM-Wiki-Overview]]
  • [[Projects/LLM-Wiki-Setup]]
  • [[Areas/Career]]