karpathy-llm-wiki

A reusable skill for building Karpathy-style LLM wikis with Claude Code, Cursor, Codex, and other Agent Skills tools.

karpathy-llm-wiki packages Karpathy's LLM Wiki idea into one installable Agent Skills skill. Your coding agent ingests sources into raw/, compiles durable knowledge pages into wiki/, answers questions with citations, and lints the wiki for consistency.

What Is an LLM Wiki?

An LLM wiki is a knowledge system where the LLM maintains structured wiki pages instead of re-searching raw documents on every question. New sources are compiled into durable markdown pages, cross-references are updated over time, and answers cite the wiki pages that already contain the synthesized knowledge.

This skill gives you three operations:

See SKILL.md for the full skill specification.

LLM Wiki vs RAG

This skill is optimized for the wiki model: knowledge that improves over time instead of re-deriving relationships on every query.

Usage Stats

Based on a production knowledge base maintained daily since April 2026:

• 94 wiki articles across 13 topic directories
• 99 source materials ingested
• 87 operation log entries in the last 7 days

See examples/ for sample wiki pages, source files, and operation logs.

Install

npx add-skill Astro-Han/karpathy-llm-wiki

Works with any tool that supports the Agent Skills standard.

Quick Start

1. Ingest your first source

Give the skill a URL, a file, or pasted text:

"Ingest this article: https://example.com/attention-is-all-you-need"

The skill stores the source in raw/, then compiles or updates the right knowledge pages in wiki/.

2. Ask your wiki a question

"What do I know about attention mechanisms?"

The skill searches the wiki and answers with citations linking back to your markdown pages.

3. Keep the wiki healthy

"Lint my wiki"

Checks for broken links, missing index entries, stale cross-references, and related issues.

How the Workflow Works

The core idea from Karpathy: the LLM maintains the wiki while the human focuses on choosing sources and asking good questions.

your-project/
├── raw/            ← Immutable source material
│   └── topic/
│       └── 2026-04-03-source-article.md
├── wiki/           ← Compiled knowledge pages maintained by the LLM
│   ├── topic/
│   │   └── concept-name.md
│   ├── index.md    ← Global table of contents
│   └── log.md      ← Append-only operation log

Each new source can update multiple pages, strengthen cross-references, and record contradictions. That is what makes the wiki compound over time.

Tool Compatibility

This skill follows the agentskills.io open standard:

FAQ

What is the difference between an LLM wiki and a personal wiki?

An LLM wiki is maintained by the model. It updates summaries, cross-links, index entries, and contradictions as new material arrives. A normal personal wiki depends on manual editing.

What sources can I ingest?

Web pages, papers, blog posts, PDFs, markdown files, text files, and pasted text. The skill converts everything into markdown under raw/ and compiles it into wiki/.

Is this production-ready?

The workflow is based on a real knowledge base with 94 articles and 99 sources maintained daily since April 2026. The repo includes examples, templates, and a design spec.

Design Boundaries

Deliberately not built, after three months of production logs and a survey of the ecosystem (LLM Wiki v2, llm-wiki-compiler, OKF, agent-memory literature):

• Source-hash freshness tracking — raw/ is immutable, so hashes guard against events that cannot happen. Genuinely new information arrives as new sources through normal ingest.
• Persisted line-number citations — every observed fidelity error was "value absent from the source", which a whole-file grep catches. Anchors only disambiguate a failure mode that has not occurred, and the annotation friction makes agents skip the rule.
• Numeric confidence or quality scores — false precision with no calibration behind it. Evidence strength belongs in the prose.
• Per-article review dates — nobody can predict at compile time how fast a domain moves. Maintenance is driven by whole-wiki lint, not per-page timers.
• Access-based decay — frequently asked is not the same as true.
• Retract / bad-source machinery — has not happened yet. Handle it manually until it does.
• Automatic hooks and scheduled runs — those belong to the agent harness, not a tool-agnostic skill.
• Vector or graph search — at 50K–100K tokens of curated wiki, grep and read are more reliable. Add search tooling only when recall measurably degrades.
• Typed relationship ontologies — link semantics live in the prose around the link.
• OKF conformance — the spec is a v0.1 draft with a minimal tooling ecosystem. Tracked; will be revisited.
• MCP servers, UIs, output subsystems — outside the boundary of a tool-agnostic skill.

Inspired By

Unofficial community implementation of the workflow from Karpathy's LLM Wiki idea. The value here is the reusable workflow, prompt structure, and battle-tested knowledge-compilation rules.

See also: lucasastorian/llmwiki, atomicmemory/llm-wiki-compiler. We are tracking Google's Open Knowledge Format draft and will evaluate compatibility once the spec and tooling mature.

License

MIT