A tweet by @ridark_eth has pulled 15,000 likes and 1,600 retweets in under 24 hours, amplifying an idea from Andrej Karpathy — OpenAI co-founder, former Tesla AI director — that originally racked up 16 million views: stop using AI to write code, and start using it to build a second brain.
The pitch is simple. You install Obsidian, the free markdown-based note app, create a vault, and open it in Claude Code, Anthropic’s terminal-based coding agent. You paste in Karpathy’s wiki idea file — a prompt that tells Claude how to structure knowledge — and tell it to build.
Claude creates three folders: raw for source material, wiki for its generated pages, and a CLAUDE.md instruction file that governs how the system behaves. From there, you drop in anything — an article, a PDF, a transcript, a research paper — and say “ingest this.” Claude reads it, extracts key concepts, cross-links them to existing pages, and files it into a living wiki of everything you know.
The more you feed it, the smarter it gets. It compounds like interest.
Why this landed
The idea resonates because it solves a real problem: every AI conversation starts from zero. You paste context, the model reads it, you have a conversation, and then it’s gone. Next time, you start over. A wiki built and maintained by AI means you never start from a blank chat again. You ask questions across everything you’ve ever fed it, and it answers with citations from your own knowledge base.
This isn’t a new concept — personal knowledge management tools like Obsidian, Notion, and Roam Research have been around for years. What’s new is the AI layer. Instead of manually linking notes and writing summaries, Claude Code does the grunt work. You provide raw material; it builds the structure.
The NZ angle
For New Zealand businesses and professionals, this is a practical AI use case that doesn’t require a cloud subscription or a computer science degree. Obsidian is free, runs locally, and stores everything as plain markdown files — your data never leaves your machine unless you choose to sync it. Claude Code requires an Anthropic API key (pay-as-you-go, roughly $3–15 per million tokens depending on the model), but the ingestion step is cheap — reading and summarising a document costs cents, not dollars.
We’ve been advocating this approach at Singularity.Kiwi. Our own editorial pipeline runs on a similar principle — a wiki of entities, topics, and source references that grows with every article we publish. The Lobster Farm, our free ebook on building local AI agents, covers the same pattern: point an AI agent at your files, let it organise and cross-reference, and get a knowledge base that compounds over time.
The catch
The idea is sound, but the execution has friction. Claude Code is a terminal tool — there’s no GUI. You need to be comfortable with a command line, or at least willing to follow a step-by-step guide. Obsidian’s vault structure requires some upfront thought. And the “ingest this” step works well on text, but struggles with scanned PDFs, images, and audio without additional tooling.
There’s also the question of lock-in. Your wiki lives as markdown files, which is about as open as it gets — but the AI layer depends on Claude Code and Anthropic’s API. If Anthropic changes pricing, deprecates a model, or goes down (as it did in a widely reported outage earlier this month), your wiki doesn’t break, but your ability to maintain it does.
The solution is the same one we recommend for everything AI: keep your data local, keep your tools swappable, and don’t build a system that only works with one provider.
How to try it
- Download Obsidian (free, Mac/Windows/Linux)
- Install Claude Code (requires an Anthropic API key)
- Create an Obsidian vault and open it in Claude Code
- Tell Claude: “Build me a personal wiki. Create a
rawfolder for sources, awikifolder for pages, and aCLAUDE.mdthat explains how to ingest and cross-link.” - Drop a file into
rawand say “ingest this” - Ask questions across everything you’ve fed it
Five minutes to set up. The payoff compounds.
