Karpathy Wiki Protocol

Karpathy Wiki Protocol — Claude Code / Codex / Cursor Skill for LLM-Maintained Knowledge Bases

**The only Karpathy wiki implementation born from a production codebase (139 pages, 594 wikilinks, 425 audited gaps), not a weekend experiment.**

---

**Skill for:** Claude Code | Codex | Cursor | Any AI Agent with filesystem access

**Category:** AI Agents | LLM Prompt Engineering | Knowledge Management | Wiki Maintenance

---

Most Karpathy wiki implementations are demos. They ingest 5 articles, generate pretty pages, and stop.

This repo is different. It's the **operational ruleset** — extracted from a real project where AI agents maintain a 139-page wiki across 22 domains, with a verification gate that catches every missed update. Every rule exists because an agent violated it in production and we had to write the rule to stop it from happening again.

---

What This Is

A complete, copy-paste-ready **AI agent skill** for making LLM agents (Claude Code, Codex, Cursor, any agent with filesystem access) build and maintain a **Karpathy-style wiki**. It's not a tool. It's not a CLI. It's a **set of battle-tested rules** you drop into your `CLAUDE.md`, `AGENTS.md`, or `.claude/` directory and your agent immediately knows how to maintain your knowledge base.

What This Is NOT

• ❌ Not a RAG pipeline
• ❌ Not an Obsidian plugin
• ❌ Not a CLI tool
• ❌ Not a SaaS product

It's a **prompt/skill**. A very battle-tested one.

---

The Problem It Solves

Every new AI agent session on your codebase starts from zero. It re-reads your docs, re-derives relationships, re-discovers gaps. **Thousands of tokens burned on initialization, every single time.** Knowledge doesn't accumulate.

The Karpathy wiki pattern fixes this: a **compiled knowledge layer** (markdown files in `wiki/`) that agents read FIRST instead of your raw docs. But the pattern only works if the wiki stays maintained. **That's where every implementation fails** — the wiki rots because nobody defined the m