Breaking

Karpathy's 'LLM Wiki' and Obsidian's Claude Code Skills Fuel a Self-Maintaining Note System

July 15, 2026 at 18:46 EDT

AI Agents
Software Dev & Coding
Open Source

A pattern published by Andrej Karpathy and a companion toolkit from Obsidian's chief executive have converged into a widely shared blueprint for a knowledge base that reads, updates and audits itself using large language models. The combination treats a personal collection of notes less like a static archive and more like a codebase that an AI agent continuously grooms.

The AI Second Brain

A note vault that reads, updates and audits itself

Andrej Karpathy's LLM Wiki pattern plus Obsidian CEO Steph Ango's Agent Skills for Claude Code have merged into a widely shared blueprint: treat a personal knowledge base like a codebase an AI agent grooms continuously — all in plain Markdown, no vector database, no RAG.

core operations: ingest · query · lint

Agent Skills for Obsidian's native formats

40K

words of writing fed into the demo system

10K+

GitHub stars across the related repos

The self-maintaining loop

Ingest

Drop raw material into raw-sources/ → agent builds structured Markdown pages in wiki/

→

Query

Answers questions from your notes, in your own words — no RAG plumbing

→

Lint

Weekly pass surfaces contradictions and prunes stale or orphaned entries

A different architecture

Typical AI note tools
(Notion AI, Mem, Reflect)

Subscription

Vector search

RAG pipeline

Plugin button

Karpathy + Obsidian
agent pattern

Plain Markdown

Agent-driven

no vector DB

no subscription

Enthusiasts say

Vault reads, updates and audits itself
Answers in your own words, no RAG
Inverts the "graveyard of orphaned notes" failure mode
Obsidian's own CEO shipping the skills lends credibility

Skeptics caution

Initial setup takes about an hour
Consistency depends on the underlying model's quality
Long-term reliability still unproven — very new
No official benchmarks published

The open question: will it stay tied to Claude Code, or spread to other agents — and does it hold up at scale beyond the early adopters?

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.