On April 2nd, Andrej Karpathy posted a thread about using LLMs to build personal knowledge bases. 20.9 million people saw it. Six weeks later it is still spreading, and this past week X has been full of people saying the combination of Claude and Obsidian is so powerful it should not be allowed, claiming one properly built vault replaces an operations director, an analyst, and a content strategist combined. I built it last week. The graph view in Obsidian already has my projects, my blog posts, and everything I am working on all linked together and it looks genuinely impressive. Then I asked one question about token costs and the whole conversation changed.

Who Is Andrej Karpathy

Andrej Karpathy is one of the most respected names in AI. He was a founding member and former Director of AI at Tesla, where he led the team building the neural networks behind Autopilot. Before that he was a research scientist at OpenAI. He now runs his own AI education platform and is widely followed for breaking down complex AI concepts in ways that actually make sense. When Karpathy posts about something he is experimenting with, the AI community pays attention. That is why a single tweet about personal knowledge bases hit 20.9 million views in six weeks.

What Is Karpathy’s LLM Wiki and Why Is It Everywhere

The Karpathy LLM wiki idea is deceptively simple. You create a folder with two subfolders: raw for your source material and wiki where the LLM organises everything.

You drop in articles, blog posts, notes, PDFs, whatever you want remembered. Then you point Claude Code at the folder and tell it to ingest. It reads every source, writes structured wiki pages, links concepts together, and maintains an index and a log automatically. You never write the wiki yourself. The LLM owns that layer entirely.

Karpathy’s own wiki on a recent research topic had around 100 articles and 400,000 words. He uses it to ask complex questions, generate slide decks, run health checks, and file his own explorations back in so every query makes the knowledge base richer over time. The idea behind the Karpathy LLM wiki is that knowledge compounds instead of disappearing into chat history the way it normally does.

Obsidian sits alongside it as the viewer. You browse the wiki, follow the wikilinks between pages, and watch the graph view fill up as your knowledge base grows.

How to Set Up the Karpathy LLM Wiki With Claude Code

The setup is straightforward. You need Claude Code and Obsidian installed. Create a folder for your wiki with raw and wiki subfolders inside it. Then open Claude Code in your terminal, navigate into that folder, and paste in Karpathy’s gist, which you can find at his GitHub. The gist is the starting prompt that tells Claude Code the entire pattern. It is intentionally abstract so the LLM can adapt it to your specific needs.

Here is the prompt to add immediately after pasting the gist:

"I want to set up an LLM wiki for [your name or project]. It will cover [describe your areas: business and projects, content research, technical builds, etc.]. The wiki folder will be at [your folder path] with raw and wiki subfolders already created. Before doing anything, ask me any clarifying questions you need. Then create a CLAUDE.md schema file tailored to my setup and scaffold the full folder structure."

Claude Code will ask you clarifying questions about your projects, key people, preferred conventions, and how involved you want to be during ingestion. Answer them honestly. It then generates a CLAUDE.md schema file that governs how everything gets filed going forward. Every future ingest session reads that file first.

I used the Obsidian Web Clipper browser extension to save six of my recent blog posts into the raw folder, ran the first ingest prompt, and watched Claude Code process each one, write summary pages, update entity pages, cross-reference concepts, and log everything. The graph view filled in. Pages started linking to each other. It worked exactly as advertised.

The Part the Viral Posts Are Not Talking About

Here is where the Karpathy LLM wiki conversation needs to be more honest.

The heavy token usage in this system is the wiki writing during ingestion. Every time Claude Code ingests a new source it reads the full file, checks existing wiki pages for context, then writes or updates ten to fifteen pages in one pass. That is a significant amount of reading and writing per source document. At any real publishing or research volume, the token cost accumulates fast.

I am on Claude Pro, not Max. That matters. I asked whether a local LLM like Gemma 4 running through Ollama could handle ingestion instead, bringing the cost to zero. Claude Code itself pushed back honestly. The ingest workflow requires reading the schema, understanding the full wiki structure, writing coherent markdown with correct wikilinks, and updating multiple files in a single pass. Smaller local models tend to produce inconsistent output, miss cross-references, and quietly accumulate errors that are hard to spot until a health check runs.

The more realistic split is using a local model for preprocessing work like converting files to clean markdown or stripping WordPress XML into individual posts, and letting Claude Code handle the actual wiki writing. That saves some tokens on the mechanical work but the expensive step still happens in Claude Code.

What Actually Makes Sense for Running the Karpathy LLM Wiki

For content research at high publishing volume, the token cost of ingesting every article you read would outpace the benefit quickly.

After working through the cost question properly, the practical answer is this: the Karpathy LLM wiki pattern is not a system you feed constantly. It is a system you feed deliberately.

For content research at high publishing volume, the token cost of ingesting every article you read would outpace the benefit quickly. That use case does not make financial sense on a Pro plan unless you are very selective about what gets ingested.

Where the Karpathy LLM wiki genuinely earns its keep at low token cost is project and personal knowledge memory. Your projects change slowly. Technical decisions, stack choices, lessons learned, notes from experiments. You ingest once when something significant happens, not constantly. At that pace, maybe four to eight ingests a month, a Pro plan handles it comfortably with plenty left over for actual writing and building.

The wiki becomes most valuable not as a content firehose but as a persistent knowledge brain. When you context switch between projects after weeks away, when you want to query everything you know about a specific tool or decision, when you need to find a connection between two things you read six months apart, the wiki surfaces it without you having to reconstruct anything from scratch.

The Graph View Is the Thing

One thing the viral posts do get right: looking at the Obsidian graph view after the first ingest is genuinely satisfying. Your projects, your research, your content, all connected and browsable. It is not magic but it is a real thing that did not exist before, and it compounds every time you add something new.

The people saying the Karpathy LLM wiki replaces an entire team are overselling it. The people saying it is too complex to bother with are underselling it. The honest version is somewhere more interesting: it is the first knowledge system that actually gets easier to maintain over time instead of harder, because the LLM does the filing and you just have to keep feeding it the right things.

FAQ

How much does it cost to run Karpathy’s LLM wiki with Claude Code? It depends on how frequently you ingest new sources. Each ingest touches ten to fifteen wiki pages per source document, which uses a meaningful amount of tokens. At low volume, four to eight ingests a month, a Claude Pro plan handles it without issue. At higher research or publishing volume, ingesting everything you read would accumulate cost quickly. Being selective about what goes into the raw folder is the main lever you have on cost.

Can I use a free local AI like Ollama instead of Claude Code for the Karpathy LLM wiki? For preprocessing work like converting files to clean markdown, yes. For the actual wiki writing, the quality drops noticeably with smaller local models. The ingest task requires consistent instruction following, accurate wikilinks, and coherent cross-referencing across multiple files. Smaller models struggle with exactly that. The practical split is local model for file cleanup, Claude Code for the wiki writing itself.

Is Obsidian necessary to set up a Karpathy LLM wiki? No. The wiki is just a folder of markdown files. Obsidian is the viewer, not the engine. Claude Code is the engine. You could use VS Code or any text editor and the system works the same way. Obsidian earns its place through the graph view and the Web Clipper browser extension, which makes saving web articles to your raw folder a one-click habit. But it is optional, not required.

I have been running this for just over a week and the honest verdict is that it works best when you treat it as a slow-burning knowledge asset rather than a daily content tool. The graph is growing. The connections are real. And the token question is one worth thinking through before you start rather than after. If you want to follow how this develops over the coming months, the August Wheel newsletter is where I share the unfiltered version of experiments like this one. Sign up at newsletter.augustwheel.com.