llm-wikis

Multi-wiki sync observability for the Karpathy LLM Wiki pattern.

What is this?

Karpathy's LLM Wiki pattern: instead of RAG querying raw chunks at runtime, an AI agent reads your source files and incrementally compiles them into a structured, persistent, interlinked Markdown wiki. It solves RAG's statelessness and chunk-boundary noise — the AI owns the overhead, you own the sourcing and querying.

llm-wikis adds the sync layer that the pattern ships without:

• Multiple independent wikis per project — each covering a different knowledge domain, under a single /wikis command surface
• Sync visibility — see exactly what's changed, added, deleted, or renamed before any tokens are spent
• Rename detection — file moves resolved from hash comparison, zero LLM tokens
• Deferred ingestion — flag files for later with --later, no surprise token costs
• Provenance tracking — bidirectional manifest linking every raw/ file to the wiki/ pages it produced
• Wiki integrity auditing — detect and resolve manually-edited wiki files before they silently diverge

Works with Claude Code, Gemini CLI, and Codex via a neutral WIKI.md schema (not hardcoded to CLAUDE.md).

Install

npm install -g llm-wikis

This installs three CLI scripts and places the skill file at ~/.claude/skills/wikis/.

Slash commands

Use /wikis directly in Claude Code (or any supported LLM CLI):

| Command | Description | |---|---| | /wikis init <name> | Create a new wiki at wikis/<name>/ | | /wikis use <name> | Set the active wiki | | /wikis list | List all wikis with sync state | | /wikis status | Show what's changed since last ingest | | /wikis ingest | Ingest unsynced raw files into wiki pages | | /wikis ingest --later | Mark files for later — zero token usage | | /wikis trace <file> | Show bidirectional provenance for any file | | /wikis report | Generate a filterable HTML provenance report | | /wikis recover manifest | Restore manifest from backup | | /wikis rebuild-manifest | Re-derive manifest from disk scan | | /wikis recover schema | Reset WIKI.md to the default template |

CLI scripts

Three deterministic scripts are available directly (used internally by the skill, but callable standalone):

# List all wikis in a project
wikis scan --list <project-root>

# Show sync status for a wiki
wikis scan --status <wiki-path>

# Record a raw → wiki ingestion in the manifest
wikis record <wiki-path> <raw-file> <wiki-page>

# Generate a static HTML report
wikis report <wiki-path> <output.html>

Project structure

Running /wikis init creates:

project/
  wikis/
    <name>/
      raw/          ← drop source files here (PDFs, notes, docs)
      wiki/         ← compiled wiki pages live here
      WIKI.md       ← schema: domain conventions for the LLM
      .wiki-manifest.json
      .wiki-manifest.backup.json
  .wiki-context     ← active wiki name (set by /wikis use)

Multiple wikis are fully independent — different raw sources, different wiki pages, separate manifests.

Manifest

Each wiki has a bidirectional manifest tracking every raw/ → wiki/ relationship:

{
  "version": "1.0",
  "wiki_name": "auditor",
  "last_ingestion": "2026-05-30T12:00:00Z",
  "raw_to_wiki": {
    "raw/paper.pdf": {
      "hash": "sha256...",
      "ingested": 1748600000000,
      "deferred": false,
      "wiki_pages": ["wiki/concept-a.md"]
    }
  },
  "wiki_to_raw": {
    "wiki/concept-a.md": {
      "hash": "sha256...",
      "sources": ["raw/paper.pdf"]
    }
  }
}

The LLM never writes the manifest directly — only the deterministic scripts do, atomically, with a backup before every write.

License

MIT