Back
Personal Brain · pattern + build-along

Wiki + Journal + CRM, all compiled by Claude Code

Karpathy described the LLM-compiled-wiki pattern. Matt Wolfe extended it with Journal + CRM and built it on stream. We use Claude Code (not Codex) and wire the same vault you set up in Step 5. Same Obsidian frontend; same agents.md control file; ~30 min to working.

Attribution. The canonical spec is Andrej Karpathy's LLM Wiki gist — read it first; everything else here is implementation. The Wiki + Journal + CRM extension and the step-by-step build is from Matt Wolfe's video. We adapt the build to Claude Code and the same vault you set up in Step 5.

Three layers (Karpathy)

  • raw/ — immutable sources. Ground truth.
  • wiki/ — LLM-owned MD pages with cross-references.
  • schema.md — config doc that turns a generic chatbot into a disciplined maintainer.

Three operations

  • Ingestion — process new sources; extract key info; maintain cross-refs.
  • Querying — wiki first; file valuable analyses back as new pages.
  • Maintenance — periodic health checks: contradictions, orphans, gaps.

Two navigation files

  • index.md — table of contents by category.
  • log.md — chronological record of ingests, queries, maintenance.
  • These are always present.

Three pillars

Wiki / Knowledge Base

Articles, YouTube transcripts, podcast notes, tweets — anything you find on the web. Compiled by an LLM into entity / concept / tools / themes pages with backlinks. Auto-linked, zettelkasten-style.

Lives at corpus/ (company SOPs) + personal-brain/wiki/ (web saves). Already part of the framework's Step 5.

Journal (grounded)

Write 'journal' to start a chat with Claude Code. Replies are grounded in your wiki, past journal entries, AND your CRM — not generic LLM advice. Files saved as journal/<date>-<slug>.md.

Optional Step 5 sub-action. The journal/ folder + agents.md rules generate themselves the first time you ask.

CRM (people)

One Markdown file per person under CRM/. Tell Claude Code who you met + where + when + any notes. Later: 'where did I meet Matthew Berman?' returns the answer instantly.

Optional Step 5 sub-action. Same vault, same pattern, separate folder.

7-step build-along (Claude Code)

Assumes you have Obsidian + Claude Code installed (covered in Step 4 + Step 5). Same vault you already set up.

1

Open Obsidian and create a fresh vault — or reuse your bettroi-vault

Either path works. The Personal Brain folders (raw/, wiki/, journal/, CRM/) coexist alongside your company SOPs in corpus/ — they don't conflict.

2

Open Claude Code in the vault folder

Claude Code runs against the current directory. Once it's pointed at the vault, every command operates on those files.

Start the Claude Code session
cd ~/BeBrain/<your-vault>
claude
3

Ask Claude to scaffold the architecture

One prompt. Claude builds raw/, raw/processed/, wiki/, agents.md, index.md, log.md following Karpathy's spec.

Paste this prompt verbatim
Build out the wiki architecture in this Obsidian vault following Karpathy's
LLM wiki pattern:

- raw/                — inbox folder (Web Clipper drops files here)
- raw/processed/      — files that have already been ingested
- wiki/               — AI-compiled .md pages (entities, concepts, tools, themes), with [[backlinks]]
- agents.md           — operations spec:
    INGEST: read files from raw/, write/update wiki + index + log,
            then move source to raw/processed/. For YouTube clips,
            extract the channel name and add it to the source's frontmatter.
    QUERY:  search the vault index first; ground every answer in pages
            you find there; if a relevant wiki page is missing, draft one
            and add it to the index.
- index.md            — table of contents for every wiki page
- log.md              — append-only row per agent action
4

Install the Obsidian Web Clipper Chrome extension

One-click capture of any article or YouTube transcript → drops it into your vault's raw/ folder. Configure it to point at your vault, set the destination folder to raw/, tag with web-clip.

Get the Web Clipper
5

Process raw/ for the first time

Clip 3-5 things you've recently watched / read. Then ask Claude to compile.

In your Claude Code session
Process any new files in raw/ following agents.md. Move processed files to raw/processed/.

Open Obsidian to watch wiki pages appear, the index update, and log entries pile up.

6

Add Journal + CRM (optional but recommended)

Extends agents.md with two more rules: chats starting with journal get saved as dated entries; details about people get filed into CRM/<name>.md.

Tell Claude Code once
Update agents.md with two new operations:

JOURNAL: when a chat starts with the word "journal", save the entire
  conversation as journal/<YYYY-MM-DD>-<slug>.md. Reply must be grounded in
  (a) wiki pages, (b) past journal entries in journal/, (c) CRM records.
  Append to journal/index.md with date + 1-line summary.

CRM: when I say "add to CRM" followed by a person's details, create or
  update CRM/<person-name>.md with their summary. Maintain CRM/index.md
  with name + 1-line bio, alphabetical order.
7

Schedule it: hourly auto-ingest

Same LaunchAgent we use for the Obsidian Git auto-sync. One-line script that runs Claude Code in non-interactive mode against the vault. Anything new in raw/ gets compiled without you touching the terminal.

Add to your hourly LaunchAgent
#!/usr/bin/env bash
cd "$HOME/BeBrain/<your-vault>" || exit 1
claude --no-tty "Process any new files in raw/. Update wiki, index, log per agents.md.
After processing, commit and push to origin/main with message:
\"vault: AI sync $(date '+%F %T')\""
What you've just built

A self-improving second brain

  • Save anything from the web with the Web Clipper → it lands in raw/ → Claude Code compiles into wiki pages with auto-backlinks.
  • Ask Claude anything and it grounds the answer in everything you've ever saved — not generic LLM training data.
  • Journal daily and Claude responds with insights pulled from past wiki, past journal entries, AND your CRM.
  • Track everyone you meet in the CRM. Ask "where did I meet X?" → instant answer.
  • Hourly automation — you never run the processing manually. Wiki gets smarter while you sleep.
  • Backed up to GitHub on every commit. Switchable to any other LLM provider — your vault is just Markdown files.

Watch the original build

Build Your AI Second Brain (Obsidian + LLM, Karpathy pattern) 35 min
Open on YouTube

Matt builds this in Codex. Everywhere he says "Codex", you can substitute Claude Code with the same prompt — same files, same architecture, same outcome.

Apply it now
Step 5 — Set up the vault
Open Step 5
Library
AI Skills harvested by ironbark
Browse skills
Methodology
The Be AI-First Framework
Read the framework