Obsidian + Claude Code: drugi mózg AI w jednym prompcie (metoda Karpathy'ego)
You have Obsidian or Notion. With 300, 500, maybe 1000 notes inside. You last opened it two months ago. With guilt. Welcome to the club.
On April 4, 2026, Andrej Karpathy (former Director of AI at Tesla, OpenAI co-founder) published a short gist on GitHub — LLM Wiki. It's not another tool. It's a paradigm flip: your second brain shouldn't be written by you. It should be written by the LLM. You just drop sources and ask questions.
In this article I'll show:
- Why the classic second brain almost always fails
- How Karpathy flips the equation — 3 layers, 3 operations
- A ready-to-paste prompt for Claude Code that sets up the whole system in 15 minutes
- Bonus: how Obsidian Web Clipper turns any article on the web into a vault asset
No prep, no $500 course required. Let's go.
Why the classic second brain fails
In 2022 Tiago Forte published "Building a Second Brain." He sold 400,000 copies by end of 2025. A whole industry grew around it — courses, tools (Notion, Obsidian, Roam), productivity influencers. And most people still abandon the system within a few months.
Why? Because the classic second brain demands from you:
- Classification of notes (PARA: Projects, Areas, Resources, Archives)
- Tagging so you can find them later
- Linking between notes
- Updates when something changes in 5 places at once
- Cleaning dead notes and broken links
That's administrative work, not thinking. After a month you have a choice: run your business or catalog notes. Everyone picks the business. The system dies.
Karpathy puts it directly: "Most people use LLMs with documents like RAG — you upload files, the LLM searches for context from scratch on every question. Knowledge doesn't accumulate." That's the core insight.
What Karpathy changes
Karpathy didn't try to build a better notebook. He asked a different question: what would a knowledge base look like if the primary reader were an LLM, not a human?
The consequence is simple and radical. A second brain consists of 3 layers on which you perform 3 operations. Everything else — classification, tags, links, audit — happens automatically, because the LLM does it.
Quote from Karpathy's gist:
The wiki is a persistent, compounding artifact. Cross-references exist permanently. Contradictions get flagged immediately. Synthesis improves with every source added.
This is a game-changer for anyone who generates a ton of context — research, client calls, strategic decisions.
3-layer architecture
secondbrain/
├── _CLAUDE.md # vault's constitution
├── _index.md # catalog (read first)
├── _log.md # chronological action log
├── sources/ # LAYER 1: raw inputs
│ ├── articles/
│ ├── youtube/
│ ├── x/
│ └── podcasts/
├── inbox/ # fleeting thoughts
└── wiki/ # LAYER 2: written by the LLM
├── concepts/
├── projects/
├── people/
└── synthesis/
Layer 1 — sources/ is immutable. Whatever lands here, you don't edit it. Source of truth. A year from now it lets you verify where Claude got a particular claim.
Layer 2 — wiki/ is dynamic. Only the LLM writes here. You never create these files manually. If something looks wrong — you fix the prompts, not the files.
Layer 3 — _CLAUDE.md is the constitution. It tells the LLM the rules: folder map, what note types to name how, when to auto-save, when to ask. One file replaces months of searching for "the right method."
Karpathy also adds two special files: index.md (catalog read first on navigation) and log.md (append-only log of LLM actions).
3 operations (daily routine)
Ingest — you drop a source, the LLM compiles
Read a McKinsey article on AI in manufacturing? Instead of leaving it in bookmarks, you drop the URL into Claude:
/obsidian-ingest https://mckinsey.com/...
What Claude does in 30 seconds (alone, not you):
- Fetches the article, saves the raw version to
sources/articles/
- Identifies companies, people, concepts (Siemens, Industry 4.0, MES)
- Updates 5-15 existing wiki pages — if "Siemens" already exists, it gets a new note. If it's a new concept, a new page gets created
- Logs the event in
_log.md
You pasted one URL. You got 15 updated knowledge pages.
Query — you ask, LLM answers from your knowledge
Next week you're on a client call about AI in manufacturing. Instead of googling from scratch, you ask Claude:
/obsidian-find "AI manufacturing ROI"
Claude searches wiki first, then sources, then daily notes. Returns a synthesis from your own knowledge. Valuable answers become new wiki pages. The vault grows not just from saved content, but also from your questions.
Lint — vault health audit
Once a week the LLM runs a review:
- Contradictions (3 months ago you wrote that a client pays 14k, today 32k — should we update?)
- Orphan pages (with no incoming links)
- Broken wikilinks
- Stale claims (with old dates worth verifying)
Zero manual work. The LLM doesn't forget, doesn't get bored, and touches 15 files in one pass.
Karpathy verbatim:
Humans curate sources, direct the analysis, ask good questions. The LLM does everything else.
Tutorial: one prompt, the whole system
This is where this article differs from other guides online. I won't give you 7 manual steps to click through. I'll give you one prompt that hands Claude the source materials and lets it design the implementation for your specific vault.
This is the Karpathy way: humans curate sources and ask, the LLM does the rest.
Requirements
- Claude Code installed and signed in (also works in Cursor, Codex CLI, Gemini CLI)
- Obsidian with an existing vault (or an empty one)
- Path to your vault
Prompt to paste
Open Claude Code in your terminal inside your vault directory and paste:
I want to build a second brain in my Obsidian following Andrej Karpathy's LLM Wiki pattern.
Here are the materials to research — read all of them before proposing anything:
1. Philosophy + architecture (primary source):
https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
2. Official Obsidian Skills by Steph Ango (Obsidian CEO) — syntax foundation, CLI, bases, canvas:
https://github.com/kepano/obsidian-skills
3. Karpathy pattern implementation with 32 slash commands + scheduled agents:
https://github.com/eugeniughelbur/obsidian-second-brain
4. Obsidian Web Clipper — for dropping articles from the browser:
https://obsidian.md/clipper
My vault: <PROVIDE PATH, e.g. ~/Documents/Obsidian/secondbrain>
Do this:
1. Research — read all 4 sources. If anything's missing, look it up.
2. Vault scan — see what I already have, what works, what's legacy to archive.
3. Plan — propose an implementation plan TAILORED to my vault (Karpathy 3-layer, folder structure, _CLAUDE.md governance, existing file migration). DO NOT implement without my confirmation.
4. After my OK — implementation:
- folder structure (sources/, wiki/, inbox/)
- _CLAUDE.md, _index.md, _log.md
- migration of old files into the new structure
- install skills from points 2 and 3 via git clone (to ~/.claude/skills/) — no npx, transparent file copy
5. After implementation — tell me what's next (e.g. deploy scheduled agents, first ingest test, Web Clipper setup).
Write concisely. Ask questions only when truly necessary — infer from the vault context.
Claude will read 4 sources, scan your vault, propose a plan tailored to your content — and roll it out after your OK.
Why this style works better than a prompt with hardcoded instructions:
- Claude reads primary sources, not my summary — catches nuance I might have missed
- The plan is tailored to your vault, not generic
- Self-updating — when Karpathy updates the gist or a new version of eugeniughelbur drops, Claude pulls the latest
- Blunt by default — "ask questions only when necessary" gives you 1-2 questions, not 10
Time: clean vault — 15-20 minutes. Vault with 500 files to migrate — 60-90 minutes (mostly waiting for migration proposals).
Bonus: Obsidian Web Clipper
Karpathy explicitly names one tool in his gist as part of the stack: Obsidian Web Clipper.
It's the official extension from the Obsidian team (built by Steph Ango, Obsidian's CEO). Works in Chrome, Firefox, Safari, Edge, Brave, Arc. Under the hood it uses Mozilla Readability — the same algorithm behind Firefox Reader View.
Killer features (as of 2026):
- Auto-apply templates — set rules per site (a McKinsey article goes to
sources/articles/, an X post to sources/x/)
- LLM Interpreter — the extension calls an LLM at clip time, generates automatic summaries, extracts key facts
- Highlights management — you highlight in the browser, it lands in the vault with context
- Everything local — no cloud, your vault is yours
Workflow:
- You read an article in the browser
- You click Web Clipper → it lands in
sources/articles/2026-05-15-title.md with metadata and a summary
- In Claude Code:
/obsidian-ingest sources/articles/2026-05-15-title.md
- Claude propagates knowledge to the wiki — 5-15 pages updated
Get it: obsidian.md/clipper.
What you gain when this works
After implementation (speaking from personal experience — I rolled this out yesterday):
- Decision time: instead of 20 minutes of research before each client call — 30 seconds of asking the vault
- Organizational memory: when you hire someone, they ask the vault. Onboarding shrinks from 2 weeks to hours
- Compounding knowledge — Karpathy calls it a "compounding artifact." Classic RAG starts from scratch every time. Here synthesis grows with every source
- Audit trail: every decision, client, deal — traceable history with dates and sources
- No cognitive load: you don't have to remember "where did I save that." You ask Claude, you get it
This isn't "a pretty notebook." It's an external brain that works faster than your own.
One condition for it to work
Karpathy warns in the gist: "Start small. 10 sources. Make sure ingest, query, lint feel natural. Then expand."
The first 2 weeks are a trial. Your work:
- Every article/X post/YT video you read for longer than 30 seconds →
/obsidian-ingest
- When you need info →
/obsidian-find instead of googling
After 14 days: if you use ingest daily, the system is alive. If not — disable it and rethink what you actually need.
Vault value = a function of use. Not setup quality. The most beautiful _CLAUDE.md in the world won't help if you go back to googling.
Want this for yourself?
If you run a company, an agency, or you're a solopreneur and you feel you're losing knowledge faster than you gain it — this is the solution.
Book a 30-min call. I'll show your use case, design the structure for your business, set up the system on one call.
Sources: