English | 简体中文

llm-wiki-starter

One command to scaffold an Andrej Karpathy’s LLM Wiki AI knowledge base.

Auto-installs Claude Code + Obsidian + recommended plugins (Skills & Plugins & Theme & Shortcuts), so AI can continuously build and maintain your personal knowledge system.

Compatible with Claude Code, Codex, Copilot, Gemini CLI, OpenCode, and other mainstream AI agents out of the box.

Installation

Option A — bash script

curl -fsSL https://raw.githubusercontent.com/eleven-net-cn/llm-wiki-starter/main/install.sh | bash

Windows users: Run the installer from Git Bash (recommended) or WSL2 — cmd.exe and PowerShell cannot execute bash scripts. Install Git for Windows (provides Git Bash + curl), then run git config --global core.autocrlf input to avoid bad interpreter errors. The installer auto-detects winget / Chocolatey / Scoop to fetch Obsidian, Node.js and Git.

With options:

# Only detect and install global tools (Claude Code, Obsidian, NodeJS, Agent Skills, etc.)
curl -fsSL https://raw.githubusercontent.com/eleven-net-cn/llm-wiki-starter/main/install.sh | bash -s -- --only-tools

# Skip global tools detection/installation, only create wiki knowledge base
curl -fsSL https://raw.githubusercontent.com/eleven-net-cn/llm-wiki-starter/main/install.sh | bash -s -- --only-wiki

# Skip tools and wiki creation, only configure Obsidian (plugins, theme, shortcuts) in current vault
curl -fsSL https://raw.githubusercontent.com/eleven-net-cn/llm-wiki-starter/main/install.sh | bash -s -- --only-obsidian

Option B — Skill install (for AI Agent users, or when bash has environment issues)

This produces the same result as install.sh. Agent-guided creation consumes tokens — recommended only when the bash script can’t run in your environment.

1. Install the Skill

   npx skills add eleven-net-cn/llm-wiki-starter -g
   

2. Trigger by chatting in your AI Agent:

   "create llm wiki"
   
   # or
   
   "scaffold an llm wiki"
   
   # or similar...
   

   The Agent will auto-detect installed tools and guide you through the setup.

3. Or, use the slash command in your Agent:

   /llm-wiki-starter
   

   The slash command supports all parameters, e.g.:

   # Only create wiki + Obsidian config, skip tool install
   /llm-wiki-starter --only-wiki
   
   # /llm-wiki-starter --bash    # Pipe through bash script — saves tokens
   

Options

Supported options (use as needed):

What Gets Installed

Detects what’s already on your system and only installs what’s missing.

Tools & Skills

• ✅ Claude Code — Recommended AI agent
• ✅ Node.js — Runtime for Claude Code and Skills CLI
• ✅ Obsidian — Wiki editor and visual graph viewer
• ✅ kepano/obsidian-skills — Obsidian Markdown, CLI interaction, Bases database views, web scraping (defuddle)
• ✅ axtonliu/visual-skills — Excalidraw diagrams, Mermaid charts, Obsidian Canvas, JSON Canvas
• ✅ Git — Version control (optional)

Skills are installed globally via Skills CLI, shared across agents.

Obsidian

• Plugins (17 plugins: 9 Core + 8 UX, auto-configured with wiki)

  Core plugins (required for llm-wiki functionality):

  • ✅ Claudian — Embed Claude Code / Codex / OpenCode agents in vault, sidebar chat with full agentic capabilities
  • ✅ Dataview — SQL-like queries on page frontmatter
  • ✅ Templater — Template system for new pages
  • ✅ Linter — Automatic Markdown formatting
  • ✅ Custom Sort — File explorer ordering via sortspec
  • ✅ Obsidian Git — Auto git commit/push (requires Git)
  • ✅ Tag Wrangler — Rename, merge, and manage tags
  • ✅ Strange New Worlds — Show wikilink reference counts
  • ✅ Homepage — Set a landing page on vault open

  UX plugins (enhance Obsidian editing experience):

  • ✅ Omnisearch — Fuzzy search across vault
  • ✅ Switcher++ — Quick switcher with headings navigation
  • ✅ Minimal Theme Settings — Minimal theme configuration
  • ✅ Hider — Hide UI elements for cleaner interface
  • ✅ Editing Toolbar — MS Word-like toolbar + F11 fullscreen shortcuts
  • ✅ Excalidraw — Hand-drawn style diagrams
  • ✅ Quiet Outline — Enhanced outline view
  • ✅ Open in Terminal — Open vault in terminal

• Theme

  ✅ Minimal — Clean, distraction-free theme (auto-downloaded)

• Key Shortcuts

  • Cmd+Shift+F → Omnisearch (fuzzy search)
  • Cmd+R → Quick switcher (headings)
  • Cmd+F11 → Workplace fullscreen
  • Cmd+Shift+F11 → Editor fullscreen focus

Browser Extension (recommended)

• Obsidian Web Clipper — Clip web articles directly to raw/inbox/ for LLM ingestion

Getting Started

# Open in Obsidian
cd my-wiki && open -a Obsidian .

# Start AI agent (also works with codex / copilot / gemini, etc.)
claude

Then chat with the AI:

• Ingest → Ingest this article: https://example.com/some-article
• Query → What is the relationship between X and Y?
• Lint → Run a health check on the wiki

Wiki Structure

my-wiki/
├── raw/                     # Immutable source materials (LLM read-only)
│   ├── inbox/               # Web Clipper inbox (auto-sorted on ingest)
│   ├── /            # Organized by knowledge domain
│   └── assets/              # Images, attachments
├── wiki/                    # LLM-maintained knowledge base
│   ├── /            # Domain-specific compiled pages
│   ├── concepts/            # Concept definition pages
│   ├── summaries/           # Source material summaries
│   ├── synthesis/           # Cross-cutting analysis
│   ├── archived/            # Deprecated pages
│   └── assets/excalidraw/   # Diagrams
├── canvas/                  # JSON Canvas visual maps
├── templates/               # Page templates (one per type, used by LLM)
├── AGENTS.md                # Wiki schema (single source of truth)
└── CLAUDE.md                # Claude Code config (imports AGENTS.md)

Tip: Domain directories (e.g. AI Agent/, Machine Learning/) are created automatically during your first ingest. Just tell the AI what domain your knowledge belongs to — or let it decide based on the content.

What is LLM Wiki?

LLM Wiki is a knowledge management pattern proposed by Andrej Karpathy: instead of traditional RAG that retrieves from scratch every query, the LLM incrementally builds and maintains a persistent wiki — cross-references are established automatically, contradictions are flagged, and synthesis is continuously updated. Each new source makes the wiki richer.

Suitable for: personal knowledge management, technical research, domain learning notes, team knowledge bases — any scenario where you want AI to help you accumulate and organize knowledge over time.

How it works: Claude Code serves as the AI agent that reads, writes and maintains the wiki; Obsidian serves as the visual editor and reader. You chat with the AI to ingest sources, query knowledge, and run health checks — while browsing and navigating the wiki graph in Obsidian.

Three-layer architecture: raw/ (immutable sources) → wiki/ (LLM-maintained pages) → Schema (AGENTS.md)

Three operations: Ingest (add knowledge) → Query (ask questions) → Lint (health check)

Credits

• Andrej Karpathy — LLM Wiki

License

MIT