Andrej Karpathy (안드레이 카파시), known for proposing the concept of “vibe coding” and as one of OpenAI’s co-founders, has drawn attention by sharing how to use AI to build a personal knowledge base.

His “secret,” posted recently on social media platform X (Twitter), requires no special software or database. All it takes is 3 folders and text files.

The key is simplicity. First, create 3 subfolders inside a project folder: raw, wiki and outputs.

Raw is where original materials are collected. Anything can go in, including articles, notes, screenshots, meeting minutes and research papers. There is no need to organize or rename them. Paste articles as Markdown (.md) or text (.txt) files, and save screenshots or charts as images.

Wiki is where AI reads the materials in raw and rewrites them in an organized form. Outputs stores AI-generated answers, reports and analysis results. After creating the folders, the next step is a schema file.

Create a single text file in the project folder named CLAUDE.md or AGENTS.md. The file serves as a guide that tells the AI what the knowledge base is and how it should be organized.

The contents are simple. It is enough to specify each folder’s role, rules for writing the wiki, and a list of topics of interest. For example, each topic in the wiki folder is created as a separate file, the first paragraph of each file begins with a summary, and related topics are linked to each other.

AI follows these instructions to convert materials in raw into the wiki. As raw fills up, users can ask AI to write the wiki. They can link AI coding tools such as Claude Code or Cursor to the project folder and give instructions. The AI reads all raw materials and creates topic-based wiki files. An index file is also created automatically.

Users do not edit the wiki directly. They read it and ask questions. The AI handles revisions and updates. Once the wiki accumulates more than 10 documents, it can be used in earnest. AI reads the entire wiki and answers questions such as "Tell me the 3 things I know least about in this topic" and "Compare how material A and material B differ in this concept." Saving those answers to outputs then becomes material for the next question.

As more materials are collected and more questions are asked, the quality of the knowledge base improves. Errors can also accumulate. If incorrect organization by the AI is saved in the wiki, later answers are built on those errors. The same point was raised in comments on Karpathy’s post.

In response, Karpathy recommends a monthly review. Users can ask the AI to check the entire wiki and find contradictions, claims without sources and topics that lack explanations.

Karpathy’s post drew a flood of comments recommending Obsidian plugins, but his response was clear. He said, "All you do is put Markdown files in a folder." AI does not care which app is used to open the files. What matters is the folder structure and the schema file. The point is to focus on actually collecting materials and asking questions instead of spending time configuring complex tools.