What is llm-wiki-anatomy?

This repository is not a piece of software you install. It is a written argument, in Korean, about how to think before building a wiki that an AI agent will maintain for you. The author's core claim is that a wiki is not a storage bin for notes, it is a search contract: before you write a single file, you should be able to name the questions you actually want answered. If you cannot name a question, what you are building is a graveyard of summaries with a pretty graph view attached, one you will stop opening after three weeks. The README breaks the idea into three points. First, a wiki is a search contract, not a storage system, so the questions come before the structure. Second, any knowledge search setup is really a combination of three layers: how information is compressed into a form, how pieces are linked together, and how an agent is meant to read them. Popular setups like an Obsidian vault are just one combination of those three layers, and the graph view people admire is a side effect, not the point. Third, a wiki is not complete unless the summary can point back to the original source document, since a search result that only shows a compressed summary with no path back to the source text is missing a required layer. The repo also references Google Cloud's Open Knowledge Format (OKF) v0.1, a proposed 2026 standard for organizing markdown files that AI agents can read across different tools, arguing that people had already been reinventing the same idea independently before this standard existed. Rather than prescribing one tool, the repository lays out six different recipes for putting the same underlying idea into practice: a markdown wiki in the Obsidian style, a SQLite database, an MCP server, a retrieval augmented generation setup, an RSS or static site feed, and an OKF-formatted bundle. A separate ontology guide covers five ways to structure the relationships between pieces of knowledge, depending on what kind of questions you plan to ask. To try it, you clone the repository and run Claude Code inside it, and the agent is designed to ask you for your search questions before it builds anything. The project is released under the MIT license.