AGI House — LLM Wiki Event

Public-editable collaborative wiki for the AGI House LLM Wiki event. Anyone can edit any page here — no account needed. Add notes, links, tips, your own wiki, whatever's useful.

If you're at the event and you've never shipped a wiki before, this page has everything you need to get to your first running wiki fast.

What is an "LLM wiki"?

A personal Wikipedia compiled by an LLM from your own life data — journals, notes, messages, tweets, bookmarks. Not a RAG index; a pre-compiled, hand-editable, plain-markdown knowledge base that any coding agent (Claude Code, Codex, Gemini CLI) can navigate via wikilinks.

Three-layer architecture (Karpathy):

Raw — original sources, untouched (raw/ or data/)
Wiki — LLM-compiled markdown articles with [[wikilinks]] (wiki/)
Schema — the SKILL.md / CLAUDE.md that instructs the LLM how to write and organize

Karpathy's pithy framing: "Obsidian is the IDE, the LLM is the programmer, the wiki is the codebase."

The tweets that made it viral

Andrej Karpathy — original LLM Wiki post (Nov 2025) — 12M+ views, kicked the whole thing off: https://x.com/karpathy/status/2040572272944324650
Karpathy's gist with the skill/prompt (2,100+ stars in 12 hours): https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
Farza Majeed — "Farzapedia" (the first widely-shared implementation, the reply that made Karpathy's idea concrete for most people): https://x.com/FarzaTV/status/2040563939797504467
Farza's published SKILL.md gist — the file you drop into .claude/skills/wiki/SKILL.md to get started: https://gist.github.com/farzaa/c35ac0cfbeb957788650e36aabea836d
Farza's landing page (interest gauge): https://farza.com/knowledge

Write-ups worth reading

Analytics Vidhya — technical breakdown: https://www.analyticsvidhya.com/blog/2026/04/llm-wiki-by-andrej-karpathy/
MindStudio — step-by-step with Claude Code: https://www.mindstudio.ai/blog/andrej-karpathy-llm-wiki-knowledge-base-claude-code
Toolmesh deep dive: https://www.toolmesh.ai/news/karpathy-llm-wiki-personal-knowledge-management
Medium — what is an LLM Wiki?: https://medium.com/@aristojeff/what-is-an-llm-wiki-and-why-are-people-paying-attention-to-it-b7e10617967d

Ingesting your iMessage history

All your iMessages live in a single SQLite database on every Mac:

~/Library/Messages/chat.db

Requires Full Disk Access for the terminal / app that opens it (System Settings → Privacy & Security → Full Disk Access).

Tables worth knowing:

Table	What's in it
`message`	Every message — `text`, `attributedBody` (emoji/rich text), `date` (Apple epoch), `is_from_me`, `handle_id`
`handle`	Phone numbers and emails of everyone you've ever texted
`chat`	Conversations (1:1 or group)
`chat_message_join`	Maps messages → chats

Gotchas:

date is in Apple epoch (nanoseconds since 2001-01-01 UTC). Convert with datetime(message.date/1e9 + 978307200, 'unixepoch') in SQL.
Modern messages store text in attributedBody (typedstream-encoded blob), NOT text — you have to decode the blob. The open-source scripts below already handle this.
Resolve phone numbers to names by joining against Contacts (~/Library/Application Support/AddressBook/) — also requires FDA.

Skills / scripts that do it for you

The skill file most people use at these events is Farza's wiki skill: https://gist.github.com/farzaa/c35ac0cfbeb957788650e36aabea836d — drop it into .claude/skills/wiki/SKILL.md in an empty project, put your exports in data/, then run /wiki ingest and /wiki absorb all.

Open-source iMessage ingesters you can crib from:

kothari-nikunj/llm-wiki — most complete. Has a dedicated ingest_imessage.py that reads chat.db directly (both text and attributedBody), resolves handles to contact names via Address Book, and groups by contact + day. Also supports WhatsApp, Apple Notes, tweets, bookmarks. Ships a Next.js viewer. → https://github.com/kothari-nikunj/llm-wiki
SamurAIGPT/llm-wiki-agent — drop-in agent, works with Claude Code, Codex, or Gemini CLI. Adds NetworkX + Louvain community detection for knowledge-graph visualization. → https://github.com/SamurAIGPT/llm-wiki-agent
ussumant/llm-wiki-compiler — Claude Code plugin focused on the compile step only. → https://github.com/ussumant/llm-wiki-compiler
nashsu/llm_wiki — cross-platform desktop app wrapper. → https://github.com/nashsu/llm_wiki

Minimal first run

mkdir mypedia && cd mypedia
mkdir -p .claude/skills/wiki
curl -sL https://gist.githubusercontent.com/farzaa/c35ac0cfbeb957788650e36aabea836d/raw \
  > .claude/skills/wiki/SKILL.md
mkdir data
# put an export in data/ — an iMessage SQLite copy, Day One JSON, Apple Notes HTML, whatever
claude
# > /wiki ingest
# > /wiki absorb all
# > /wiki query "who do I talk to about X?"

Farzapedia by the numbers (reference)


Total raw entries ingested	2,500
Compiled wiki articles	400
Sources	Day One diary, Apple Notes, iMessage
Architecture	Karpathy's 3-layer (raw → wiki → schema)
Agent used to query	Claude Code
Storage	Plain markdown files — no vector DB, no embeddings
Most surprising output	An auto-generated "Relationship History" article (why Farza won't make Farzapedia public)

Add your own page

This wiki is public-edit. Anyone can:

Edit this page — click Edit at the top and save.
Create a new page — go to /@jacobcole/agi-house-llmwiki/new or POST to /api/v1/wikis/jacobcole/agi-house-llmwiki/pages (no auth required — anonymous pages are marked claimable).
Link to your own LLM wiki — drop it under Attendee wikis below.

Pages you might want to add:

Your attempt, with what worked / what broke
A better ingester for a specific source (WhatsApp, Signal, Slack, Email)
Tips for the absorb / cleanup phase
Prompts you found useful
Trip reports on different models (Opus vs. Sonnet vs. GPT vs. Gemini)

Attendee wikis

Add yours! [[name]] — one-line description — URL

(add your wiki here)

FAQ

Do I need to pay for anything? No. wikihub.md is free. Claude Code free tier is enough to run a small absorb. Farza used Claude.

Will my private data leak? Depends entirely on where you run the compile. Karpathy's #2 principle is local — the files stay on your computer. If you point Claude Code at a local directory, inference happens via the API but the files never leave your machine except in prompts. If you're compiling anything sensitive, audit the SKILL.md first and keep the wiki directory out of any cloud sync.

How big does it get? Farza: 2,500 entries → 400 articles, a few MB of markdown. iMessage-only ingests can be 10K–100K messages; they compress down to dozens of articles because most chats are banal.

Can I push my wiki back to wikihub.md? Yes — git clone https://USERNAME:[email protected]/@USERNAME/SLUG.git and push. Or use the REST pages API. See /AGENTS.md on this host for the full agent setup recipe.

WikiHub home: https://wikihub.md
WikiHub agent setup: https://wikihub.md/AGENTS.md
Karpathy's broader take on personal AI: https://x.com/karpathy/status/2040572272944324650
Ankur Warikoo's "YBAAS" response (you-are-a-service, one of the better commentaries on why LLM wikis are actually a big deal): https://x.com/warikoo/status/2040595110635901255
IQ Source — "knowledge compounds or rots" (the case for curation): https://www.iqsource.ai/en/blog/karpathy-llm-wiki-knowledge-compounds-or-rots/

This page is public-editable. Hosted on WikiHub. Created for the AGI House LLM Wiki event, 2026-04-22.

Connect your MCP client to WikiHub

Want Claude Desktop, Claude Code, or ChatGPT Deep Research to read and write your WikiHub pages with your own API key? Follow the step-by-step setup guide:

WikiHub MCP Connector — Setup — install, configure, deploy

Short version: build mcp-server/ in the wikihub repo, paste a two-line JSON block into your Claude config, and wikihub_search / wikihub_create_page / 15 other tools become available inside your agent.