Build Karpathy's LLM Wiki for Your AI Stack

Your AI tools are everywhere. Skills in one project, agents in another, half-remembered configs in a third. You have lost track of what you run.

The fix is an LLM wiki. It is Andrej Karpathy's pattern: a folder of linked plain-text pages that an AI agent keeps up to date for you. Point it at your AI stack, and it maps every skill, tool and agent you run, how they connect, and what has quietly drifted.

Here is what it is and how to build one.

---

What an LLM wiki is

Most people feed raw documents to the model every time they ask a question. That is slow, and the model re-reads everything from scratch.

Karpathy's idea flips it. You maintain a wiki the AI reads. The expensive thinking, the synthesis and the linking, happens once, when knowledge goes in. Every later read is fast and rich.

The trick underneath it: the boring part of any knowledge base is the upkeep. Updating links. Spotting contradictions. Keeping pages in sync. That is exactly what an LLM is good at and humans are bad at. So you flip the work. You curate and direct. The agent does the maintenance.

You can read Karpathy's original LLM-wiki write-up for the full pattern. It is short and worth it.

---

Why your AI stack needs one

Once you run skills, agents and connectors across more than one project, you cannot hold it all in your head. You forget what you built and build it twice. You cannot prune what you cannot see.

A wiki pointed at your stack is the safeguard. It lists every skill and tool, shows how they connect and why, and flags the copies that have quietly diverged. It is the difference between a pile of tools and a stack you can grow on purpose.

This matters more the more AI you run. A stack you can scan is one you can trust, prune and improve. Working out what belongs in that stack is its own question, and DevWiz has a practical guide to picking your AI stack.

---

The three pieces

The build is three small parts, each answering a different question.

The manifest. A script scans your projects and writes a list of every skill, MCP, plugin and agent. It answers "what do I have?"

The wiki. A folder of linked markdown pages, Karpathy-style. You and your agent write them. It answers "how does it connect, and why?" This is the part that compounds.

The dashboard. A small app that reads those files and renders them in your browser. It answers "show me."

The split matters. The manifest is a flat list a script can build on its own. The wiki holds the things a script can never know: why you chose a tool, how two tools work together, what you decided and when.

---

You curate, the agent maintains

Here is the line to remember. A script finds the facts. You add the meaning. The agent keeps it current.

You bring the judgement: what matters, why you picked a tool, how two things connect. The agent handles the upkeep: updating links, spotting what has drifted, keeping every page in sync. Do it the other way round and the thing rots.

That is what makes it a wiki that grows more useful every week, instead of another folder nobody opens.

---

Build it

You do not build this by hand. The full guide has a mega prompt: paste it into Claude Code, answer a few questions, and it scaffolds every file, runs the first scan, and prints how to start the dashboard.

The skills that fill the wiki are their own build. If you are new to those, start with what agent skills are first. And the whole thing is one part of a bigger picture: the map in an AI Operating System, which ties your stack, your memory and your methods into one loop.

---

See the full guide

This is the short version. The full walkthrough has the complete code for both engine scripts and the dashboard, the wiki page types, the linking rules, the maintenance loop, and the mega prompt that builds the lot.

Read Build Your Own AI Stack for the whole build. Then, to see which part of your AI setup to fix first, run the assessment. Five minutes, scored, with a clear first move.