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.
git clone https://github.com/6eanut/llm-wiki
cd llm-wiki
./quickstart.shThat's it. The script installs the skill, initializes the wiki, and drops in demo source files (Greek mythology — optional, use --no-demo to skip).
After setup, start Claude Code and run:
/wiki-ingest .raw/greek-olympians.md
Then ask anything about the content — Claude checks the wiki automatically.
./quickstart.sh --with-hooksThis 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
Prefer manual control? Here's the three-step version:
./llm-wiki/install.sh --force # 1. Install skill
~/.claude/skills/llm-wiki/scripts/setup-project.sh ./wiki --with-hooks # 2. Init wiki
# 3. Drop source files in .raw/ and run /wiki-ingestSee the Quick Start Guide on the wiki for troubleshooting and advanced configuration.
After setup, start Claude Code in your project and ask a question:
You: "What is the relationship between Zeus and Athena?"
Claude: [reads ./wiki/.llm-wiki/index.md automatically]
[finds relevant pages]
[synthesizes answer with citations]
## Answer
Athena is Zeus's daughter, born from his head...
## Evidence
| Source Page | Key Point | Confidence |
|-------------|-----------|------------|
| [[athena]] | emerged from Zeus's forehead | high |
| [[zeus]] | Father of Athena | 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/.
- Drop source files (markdown, text) into
./.raw/ - Run
/wiki-ingest .raw/your-file.md - The file is analyzed, concepts extracted, and interlinked pages created
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
| Decision | Rationale |
|---|---|
| CLAUDE.md for rules | Always-loaded, no tool call needed. Tells Claude WHEN to use the wiki. |
| SessionStart hook for state | Dynamic wiki stats (pages, topics, pending items) injected each session. |
| LLM is the runtime | All content work done by Claude. No external language runtime needed. |
| Bash for determinism only | SHA-256 hashing, file listing, grep — correctness-critical operations only. |
| Markdown workflow files | Each command has a workflows/*.md procedure. Documentation = executable instructions. |
| Two-phase ingest | Phase 1 (analysis) writes a reviewable analysis before Phase 2 (generation) creates pages. |
| Auto-generated index | index.md regenerated on every change. Enables O(1) lookup of relevant pages. |
| SHA-256 incremental caching | .done sentinel files prevent re-ingestion. Safe to re-drop sources. |
.raw/ (sources) → wiki/ (pages) → skill (schema + workflows)
(immutable) (LLM-generated) (conventions)
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
type: concept
language: en | zh | bilingualBody: Definition → Key Properties → Examples → Related
type: articleFile: YYYY-MM-DD-{slug}.md
type: persontype: synthesis
query, based_on[], confidence: high | medium | lowFile: synth-YYYY-MM-DD-{slug}.md
- Auto-detection: CJK character ratio determines
zh/en/bilingual - Page titles:
"English / 中文"format for bilingual pages - Cross-language wikilinks:
aliasesfield provides translations for link resolution - Query matching: Prefers same-language pages, falls back across languages
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
| Typical RAG | LLM Wiki | |
|---|---|---|
| Knowledge state | Re-derived per query | Persisted, compounding |
| Cross-references | None | Bidirectional [[wikilinks]] |
| Contradictions | Undetected | Flagged with callout blocks |
| Confidence | Opaque | Explicit per-page ratings |
| Audit trail | None | based_on provenance chain |
| Query cost | Every query reads source chunks | Index-first: only 3-5 pages read |
| Proactive | No — must be invoked | Yes — CLAUDE.md + hook drives behavior |
Note: Advanced RAG implementations can include some of these features, but they're not standard. LLM Wiki provides them by design.
| Principle | Implementation |
|---|---|
| LLM is the runtime | All content work done by Claude. No external language runtime needed. |
| Bash for determinism | SHA-256, grep, file listing — correctness-critical operations only. |
| Auto-generated index | Regenerated on every change. Never hand-edited. |
| Incremental caching | SHA-256 sentinel files prevent re-work. |
| Two-phase ingest | Human checkpoint between analysis and generation. |
| Hot cache | Multi-session context bridge via SessionStop/SessionStart hooks. |
| True bilingual | language field, CJK detection, cross-language aliases. |
| Lint separation | Quick (bash, free) vs Full (LLM, thorough) — pay only when needed. |
- Pattern: Andrej Karpathy
- Implementation: Built with Claude Code
MIT — see LICENSE for full text.