llmwiki

llmwiki is Atomic Strata's open-source knowledge compiler, and this is its Agent Skill. It teaches your agent to use an llmwiki project: pull a compact, cited evidence pack from the wiki before answering, and build or update the wiki from your sources when you ask. It follows Andrej Karpathy's LLM Wiki pattern: compile knowledge once into durable pages instead of re-reading raw files for every question. What it does - Checks the wiki's size, pending work and freshness first - Pulls a bounded context pack of evidence for the task, cited to source files and line ranges - Reads a known page, searches pages or asks the wiki a question - Builds a new wiki from your notes, papers or docs, with generated pages staged for your review - Adds new sources to an existing wiki and recompiles it - Reports missing evidence instead of inventing it What llmwiki gives you - Typed pages (concept, entity, comparison, overview) with paragraph-level citations - Hybrid retrieval: semantic search, BM25 reranking and wikilink graph expansion - A local read-only viewer, lint and freshness checks, and an eval harness - An MCP server (llmwiki serve) and exports to JSON, JSON-LD, GraphML, Marp and llms.txt How to install it Run npx skills add atomicstrata/llm-wiki-compiler. The skill needs the llmwiki CLI (npm install --global llm-wiki-compiler, with Node.js 24 or later) or a connected llmwiki MCP server. It works in Claude, ChatGPT, Gemini and more agents that support Agent Skills. Good to know - Retrieving evidence works without provider credentials; compiling pages and generated answers need an LLM provider such as Anthropic, OpenAI, Ollama or another OpenAI-compatible server. - It runs locally: the MCP server is a local process, not a hosted search endpoint. - MIT licensed.