Now liveThe Skillselion MCP - thousands of ranked skills, loaded into your agent mid-task. No install.Get it →
ChristianASC5 avatar

Llm Wiki

  • Updated June 24, 2026
  • ChristianASC5/christianASC5-marketplace

llm-wiki is a Claude Code skill in the AI & Agent Building category. Persistent, compounding knowledge base for your project — maintained by Claude, inspired by Karpathy's LLM Wiki pattern.

Key points

  • llm-wiki
  • AI & Agent Building
  • AI-coding skill

Llm Wiki by the numbers

  • Data as of Jul 7, 2026 (Skillselion catalog sync)
/plugin marketplace add ChristianASC5/christianASC5-marketplace
/plugin install llm-wiki@christianASC5-marketplace

Add your badge

Show developers this skill is listed on Skillselion. Paste this into your README.

Listed on Skillselion
Last updatedJune 24, 2026
RepositoryChristianASC5/christianASC5-marketplace

What it does

Persistent, compounding knowledge base for your project — maintained by Claude, inspired by Karpathy's LLM Wiki pattern.

README.md

llm-wiki

A Claude Code plugin for building and maintaining a persistent, compounding knowledge base in your project — based on Karpathy's LLM Wiki pattern.

Instead of re-deriving answers from raw documents on every query (RAG), Claude incrementally builds and maintains a structured wiki of markdown files. Each source you add gets synthesized into the existing knowledge. Cross-references are maintained. Contradictions are flagged. The wiki gets richer over time.

Claude writes it. You curate it.


How it works

wiki/
├── CLAUDE.md        ← schema: conventions Claude follows (domain-specific, generated on setup)
├── index.md         ← content catalog, updated on every operation
├── log.md           ← append-only event log
├── raw/             ← your source documents (immutable — Claude never writes here)
└── pages/           ← Claude-maintained wiki pages

The raw/ directory holds your source documents. The pages/ directory is entirely maintained by Claude.


Getting started

1. Install the plugin

# From marketplace
/plugin llm-wiki

# Or locally
cc --plugin-dir /path/to/llm-wiki

2. Initialize a wiki in your project

/llm-wiki:create-wiki

Claude will ask about your domain and generate a tailored schema (wiki/CLAUDE.md) with page types, naming conventions, and index structure suited to your use case.

3. Add sources

Drop files into wiki/raw/, then:

ingest wiki/raw/my-article.md

The wiki-agent will read the source, discuss key takeaways with you, write a summary page, update related entity pages, and maintain the index and log.

4. Query the wiki

what does the wiki say about [topic]?

The wiki-agent reads the index, retrieves relevant pages, and synthesizes a grounded answer with citations. You can optionally file the answer back as a new wiki page.

5. Health-check the wiki

lint the wiki

The wiki-agent scans for orphan pages, contradictions, stale claims, and missing cross-references, then produces a prioritized report.


Components

Component Type Purpose
create-wiki Skill (slash command) Scaffold wiki + generate domain schema
ingest Skill Interactive source ingestion workflow
query Skill Wiki retrieval and synthesis workflow
lint Skill Wiki health audit workflow
wiki-agent Agent The actor — handles all wiki operations
load-wiki.sh Hook (SessionStart) Loads wiki schema + enables proactive mode

The session-start hook

When a project has a wiki, the SessionStart hook automatically:

  • Loads wiki/CLAUDE.md into Claude's context
  • Instructs Claude to consult the wiki proactively throughout the session

To disable for a session, disable the plugin (/plugin llm-wiki --disable).


Optional: search at scale

For wikis with 100+ pages, consider adding a local search backend:

  • knowledge-rag — MCP server with hybrid BM25 + semantic search, 100% local
  • qmd — lightweight markdown search with CLI and MCP interfaces

The wiki-agent will use Grep for search by default. Configure an MCP search server in your project's .mcp.json to upgrade to vector search without changing the plugin.


Credits

Pattern by Andrej Karpathy.

Related skills

This week in AI coding

Five minutes, every Monday - the tools, releases and tactics for developers.

unsubscribe anytime.