LLM wiki for Claude Code: the workflow

Karpathy's LLM wiki pattern was born in a terminal. A folder of Markdown, an agent that keeps it current, Claude Code doing the bookkeeping.

The pattern is right. The local folder is the part that breaks.

One machine holds the vault. One client can edit it. You switch laptops and the wiki stays behind.

Here is the same workflow with the wiki hosted in Hjarni instead. Claude Code reads and writes it over MCP, from any machine, and the notes outlive the session.

The loop

The workflow is four moves you repeat all day:

1. Claude Code reads the relevant notes for context.
2. You work.
3. Claude Code writes what it learned back as notes.
4. Next session, the context is already there.

That is the loop. Research in Claude Code, save to Hjarni, act, repeat. The wiki grows while you code.

Step 1: connect Claude Code to your wiki

Add Hjarni as an MCP server. The connect Claude via MCP guide has the exact config for Claude Code. It takes five minutes and you do it once.

After that, Claude Code can search your notes, read any note, and write new ones. No files on disk. No vault to sync.

Step 2: give it a folder to read

Point Claude Code at the part of your wiki that matters for the work. A starter structure for a developer wiki:

/Stack
  - Languages and frameworks
  - Deployment
  - Conventions

/Projects
  /Hjarni
    - Architecture
    - Decisions
    - Open questions

/Fixes
  - Errors I have hit before

Start a session with one line:

Read my Stack and the Hjarni project notes, then use them as context for this session.

Claude Code fetches the notes. It knows your framework, your deploy setup, your conventions. You stop typing the same preamble into every chat.

Step 3: let it write back

This is where the wiki maintains itself. As you work, hand the bookkeeping to Claude Code:

Save the root cause and the fix we just found to my Fixes folder.

Update the Hjarni architecture note with the decision we made about background jobs.

Create an open-questions note for the migration and link it to the architecture note.

Claude Code writes the notes, tags them, and links them. You did not open a separate app. The next time you hit that same deploy error, the fix is already a note.

Capture every session automatically

You will forget to ask. So do not rely on asking.

Hjarni supports hooks that capture Claude Code sessions for you, so the useful parts land in your wiki without a manual prompt. The Claude Code hooks guide walks through the setup.

Now the loop closes on its own. You code, the session is captured, the wiki compounds.

Why not a local vault

The local Obsidian and Claude Code setup works. It also has three kinds of friction:

• One machine. The vault is on the laptop you set it up on.
• One client. Claude Code can edit the files. The Claude app on your phone and ChatGPT cannot see them.
• Hard to share. You can hand someone a git repo. You cannot hand them a living wiki they query from their own AI.

Hosting the wiki removes all three. Same notes, folders, and tags. Available in Claude Code, in the Claude app, in ChatGPT, on your phone. If you want the local version with git history and a graph view, Karpathy's pattern is still the better fit, and that is an honest trade.

Prompts that carry the workflow

Keep these close:

Read my Stack note and the current project notes before you answer.

Save the useful parts of this session as a note in the right folder.

Find notes related to what we just did and link them together.

What did we decide last week about auth? Check my notes first.

The point is not to maintain a wiki by hand. The point is to let Claude Code maintain it with you.

---

New to the pattern? Start with how to build an LLM wiki with Claude or ChatGPT and MCP, then come back here for the daily loop.

See how Hjarni works with Claude Code.