LLM Wiki — Compounding Knowledge Base for Claude Code

A Claude Code skill that builds and maintains a persistent, interlinked wiki from your source documents. Based on Andrej Karpathy's LLM Wiki pattern.

Knowledge is compiled once and kept current, not re-derived on every query.

Quick Start

Basic Setup

# 1. Install the skill globally (one-time)
./llm-wiki/install.sh --force

# 2. Set up wiki in your project
~/.claude/skills/llm-wiki/scripts/setup-project.sh ./wiki

# 3. Start Claude Code and ask questions — Claude checks the wiki automatically!

With Session Hooks (Recommended)

For the full experience, install with --with-hooks:

~/.claude/skills/llm-wiki/scripts/setup-project.sh ./wiki --with-hooks

This enables:

• Dynamic wiki stats at startup — page count, recent changes, pending reviews injected at the start of every session
• Hot-cache for session continuity — context from your last session is bridged forward so you don't lose state between sessions

What to Expect

After setup, start Claude Code in your project and ask a question:

You: "What is RISC-V?"

Claude: [reads ./wiki/.llm-wiki/index.md automatically]
        [finds relevant pages]
        [synthesizes answer with citations]

        ## Answer
        RISC-V is an open standard instruction set architecture...

        ## Evidence
        | Source Page | Key Point | Confidence |
        |-------------|-----------|------------|
        | [[risc-v]] | ISA overview | high |

You don't need to type /wiki-query for routine questions. Claude reads CLAUDE.md at startup and follows the rule: "check the wiki before answering."

If the wiki doesn't have relevant knowledge, Claude will tell you and suggest adding source files to .raw/.

Adding Knowledge

1. Drop source files (markdown, text) into ./.raw/
2. Run /wiki-ingest .raw/your-file.md
3. The file is analyzed, concepts extracted, and interlinked pages created

Slash Commands

Architecture

How Proactive Wiki Works

Session starts
    ↓
Claude reads CLAUDE.md → "Check the wiki before answering"
    ↓
SessionStart hook runs → wiki stats, topics, pending items
    ↓
Slash commands auto-discovered from ~/.claude/commands/
    ↓
Skill auto-registered as "wiki" from ~/.claude/skills/llm-wiki/
    ↓
User asks any question
    ↓
Claude reads index.md → finds relevant pages → answers with citations

Key Architectural Decisions

Three-Layer Data Architecture

.raw/ (sources)    →    wiki/ (pages)    →    skill (schema + workflows)
  (immutable)            (LLM-generated)       (conventions)

Wiki Directory Structure

wiki/
├── .llm-wiki/
│   ├── schema.md                   Copy of WIKI_SCHEMA.md
│   ├── config.md                   User preferences
│   ├── index.md                    ★ AUTO-GENERATED — never edit by hand ★
│   ├── review.json                 {pending: [...], resolved: [...]}
│   ├── cache/
│   │   ├── hot-cache.md            Multi-session context bridge
│   │   ├── source-manifest.json    SHA-256 → source metadata
│   │   ├── state-hash.txt          Detects external modifications
│   │   └── ingests/{sha256}.done   Sentinel files (idempotent ingestion)
│   └── inbox/{sha256}-analysis.md  Phase 1 ingest analyses
├── transformer.md                  Concept page
├── 2026-04-28-weekly-notes.md      Article page
├── alan-turing.md                  Person page
└── synth-2026-04-28-riscv.md       Synthesis page

Page Types

concept — Define a term, idea, methodology, tool

type: concept
language: en | zh | bilingual

Body: Definition → Key Properties → Examples → Related

article — Notes, blog drafts, imported documents

type: article

File: YYYY-MM-DD-{slug}.md

person — Author, researcher, notable individual

type: person

synthesis — Saved query answer (the compounding mechanism)

type: synthesis
query, based_on[], confidence: high | medium | low

File: synth-YYYY-MM-DD-{slug}.md

Bilingual Support

• Auto-detection: CJK character ratio determines zh / en / bilingual
• Page titles: "English / 中文" format for bilingual pages
• Cross-language wikilinks: aliases field provides translations for link resolution
• Query matching: Prefers same-language pages, falls back across languages

Skill File Map

llm-wiki/
├── SKILL.md                         Skill manifest with proactive usage rules
├── WIKI.md                          CLAUDE.md template (copied to project root)
├── WIKI_SCHEMA.md                   Page type definitions & conventions
├── install.sh                       Global installation (one-time)
├── commands/                        Auto-discovered slash commands
│   ├── wiki-ingest.md               /wiki-ingest
│   ├── wiki-query.md                /wiki-query
│   ├── wiki-lint.md                 /wiki-lint
│   ├── wiki-save.md                 /wiki-save
│   ├── wiki-graph.md                /wiki-graph
│   └── wiki-review.md               /wiki-review
├── templates/                       Page templates (article, concept, person, synthesis)
├── scripts/                         Deterministic bash operations
│   ├── setup-project.sh             ★ One-stop project setup
│   ├── init-wiki.sh                 Bootstrap new wiki directory
│   ├── hash-files.sh                SHA-256 hash source files
│   ├── check-stale.sh               Index freshness check
│   ├── find-orphans.sh              Pages with zero incoming links
│   ├── validate-frontmatter.sh      Required field validation
│   └── find-broken-links.sh         Dead wikilink detection
├── workflows/                       Deep workflow procedures (read by the skill)
│   ├── ingest.md                    Two-phase source ingestion
│   ├── query.md                     Index-first knowledge retrieval
│   ├── lint.md                      Structural + semantic health check
│   ├── save-synthesis.md            Persist answers as synthesis pages
│   ├── graph.md                     D3.js knowledge graph generation
│   └── review.md                    Review queue processing
└── hooks/                           Session lifecycle
    ├── session-start.sh             Wiki stats + PROACTIVE WIKI RULE
    └── session-stop.sh              Write hot-cache for next session

Compared to RAG

Note: Advanced RAG implementations can include some of these features, but they're not standard. LLM Wiki provides them by design.

Design Principles

Community

• Contributing Guide
• Code of Conduct
• Security Policy
• Changelog
• Support
• FAQ

Credits

• Pattern: Andrej Karpathy
• Implementation: Built with Claude Code

License

MIT — see LICENSE for full text.