deep-wiki

한국어

An LLM-managed markdown wiki for persistent knowledge accumulation — a Claude Code plugin implementation of Karpathy's LLM Wiki philosophy.

"Most people's experience with LLMs and documents looks like RAG: you upload a collection of files, the LLM retrieves relevant chunks at query time, and generates an answer. This works, but the LLM is rediscovering knowledge from scratch on every question. There's no accumulation." — Andrej Karpathy

Role in Harness Engineering

deep-wiki serves as the persistent knowledge layer in the Deep Suite ecosystem. In the Harness Engineering framework's 2×2 matrix, it operates as an Inferential Guide — providing accumulated project knowledge that shapes the agent's understanding during Phase 1 Research, replacing the need for repeated RAG queries with a compounding knowledge base.

The Idea

Instead of re-discovering knowledge each time (RAG), Claude Code incrementally builds and maintains a persistent wiki — a structured, interlinked collection of markdown files. When you add a new source, the LLM reads it, extracts key information, and integrates it into the existing wiki. The knowledge is compiled once and kept current, not re-derived on every query.

The wiki is a persistent, compounding artifact. The cross-references are already there. The contradictions have already been flagged. The synthesis already reflects everything you've read.

Architecture

Based on Karpathy's three-layer model:

Raw Sources  →  Wiki (markdown pages)  →  Schema (management rules)
    ↑                   ↑                        ↑
 wiki-ingest        pages/               wiki-schema skill

Layer	Description	Owner
Raw Sources	Immutable inputs — files, URLs, text, reports	You curate
Wiki	LLM-generated markdown pages with cross-references	LLM writes, you read
Schema	Rules governing how the wiki is structured and maintained	Co-evolved

Installation

Prerequisites

• Claude Code CLI installed and configured

Via Deep Suite marketplace (recommended)

# 1. Add the marketplace
/plugin marketplace add Sungmin-Cho/claude-deep-suite

# 2. Install the plugin
/plugin install deep-wiki@Sungmin-Cho-claude-deep-suite

Standalone

# 1. Add this repo as a marketplace
/plugin marketplace add Sungmin-Cho/claude-deep-wiki

# 2. Install
/plugin install deep-wiki@Sungmin-Cho-claude-deep-wiki

Quick Start

# 1. Initialize the wiki
/deep-wiki:wiki-setup ~/Obsidian/MyVault/wiki

# 2. Ingest sources into the wiki
/deep-wiki:wiki-ingest https://example.com/article
/deep-wiki:wiki-ingest ./document.pdf
/deep-wiki:wiki-ingest  # paste text directly

# 3. Query the wiki
/deep-wiki:wiki-query What are the rules of React hooks?

# 4. Health check
/deep-wiki:wiki-lint

Commands

Command	Description
/wiki-setup	Initialize wiki and create directory structure
/wiki-ingest	Read a source (URL, file, text) and create/update wiki pages
/wiki-query	Search the wiki and generate grounded answers; auto-files cross-page syntheses back into the wiki
/wiki-lint	Health check — schema violations, orphan pages, broken links, contradictions (also runs automatically after ingest/rebuild)
/wiki-rebuild	Regenerate index.json from page frontmatter

Operations in Detail

Ingest — Drop a new source and tell the LLM to process it. The LLM reads the source, writes summary pages, updates the index, updates relevant pages across the wiki, and appends to the log. A single source might touch multiple wiki pages. New information is merged with existing content — pages grow richer with each ingest. Auto-lint runs after every ingest to keep the wiki healthy.

Query — Ask questions against the wiki. The LLM searches for relevant pages using a three-layer strategy (index scan → content search → candidate reading) and synthesizes an answer grounded in wiki content, with citations. When a query synthesizes insights across 2+ pages, the result is automatically filed back into the wiki — the knowledge compounds.

Lint — Health-check the wiki. Looks for: schema violations, contradictions between pages, orphan pages with no inbound links, broken links, stale versions, and index drift. Optionally auto-fixes structural issues with --fix. Runs automatically after ingest and rebuild — you only need to invoke it manually for deep inspections.

Rebuild — Regenerate index.json from page frontmatter. Use when the index is out of sync or corrupted. Auto-lint runs after rebuild.

Storage Structure