How to build an LLM wiki for a codebase with Claude Code

The concept of an LLM wiki—where an AI agent maintains a folder of Markdown files detailing architecture and design decisions—was originally sketched by Andrej Karpathy. A recent end-to-end implementation tested this on itsdangerous, the Flask session signing library dating back to 2011, aiming to capture the architectural context that lives between lines of code which traditional API docs like Sphinx miss.

The foundational rule making this work is rigorous sourcing: every claim must cite a file and line at a specific commit, a CHANGES entry, or be explicitly labeled as conversation-derived. If the agent cannot source a claim, it states on the page that the information is not provided in the sources rather than inventing a plausible explanation.

Stock photo for illustration only, not from the actual event

The workflow begins by initializing the wiki via a Claude Code plugin command. The agent reads pyproject.toml, selects categories, updates .gitignore, and creates an initial commit. It then writes architecture pages for each module and dedicated design decision pages sourced directly from CHANGES.rst or the code, detailing rejected alternatives.