A GitHub repository, "Karpathy-llm-wiki-bootstrap-skill," has been released that repackages Andrej Karpathy's "LLM Wiki" pattern as an installable skill for Claude Code, OpenAI Codex and others. It lets an LLM agent compile raw sources into persistent Markdown wikis that are built and maintained automatically, and it ships with a working example.
May 6, 2026 · LLM Wiki Bootstrap · v1.0.0
The LLM Reads Once, Then Builds a Living Wiki
Based on Andrej Karpathy's "LLM Wiki" pattern, an installable agent skill lets models read raw sources and incrementally maintain a structured, cross-linked Markdown wiki — a compiled knowledge artifact, not query-time retrieval.
Install as a reusable skill
npx skills add nanzhipro/Karpathy-llm-wiki-bootstrap-skill@llm-wiki-bootstrap
Runs on Claude Code, OpenAI Codex & Copilot · global install & update via npx skills update
What makes it different from RAG
Conventional RAG
Retrieves external documents at query time and appends them to the answer. Nothing persists.
LLM Wiki
Reads sources once , then builds a compiled, reusable, cross-linked knowledge artifact maintained over time.
The two-layer structure
raw/
Immutable sources — preserved, never rewritten.
wiki/
Synthesis layer: index.md · concept-table.md · overview.md · log.md
SCHEMA.md
Defines the agent's operating contract.
Three core operations
ingest
Read sources, update pages & the concept table.
→
query
Answer via index reference + optional BM25.
→
lint
Detect contradictions, staleness & orphaned pages.
Optional BM25 search via SQLite FTS5 · supports Obsidian, editors & multiple runtimes · Chinese-language version bundled.
What works
"Impressive after two weeks of use" — a memory bank for agents
Strong at mapping concepts & relationships
Self-repair through contradiction flagging
The limits
"If not used correctly, it's no different from RAG"
Weaker for practical knowledge that evolves over time
Initial bootstrap effort & token cost on large wikis
Reliability comes from a division of labor: humans choose sources and make judgments , while the LLM handles maintenance — building cross-linked wikis from large volumes of papers and curating Obsidian vaults.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…