Building complex codebases

make an AI agent reliable on a big, real repo

This track is what comes after you have picked a coding assistant. Choosing the tool is the "Write code" track; making any of them dependable on a large, real repository is this one. The move is to stop treating the agent as a chat window and start treating your project as something that has to explain itself. You build a second, version-controlled layer next to the code — a lean rules file, reusable commands, and later skills, subagents and MCP — so that every session starts already knowing your architecture instead of guessing it. Then you run each task through the same Plan → Implement → Validate loop, and turn every mistake the agent makes into a permanent fix to the layer rather than a correction you retype. Start with "Building complex codebases" for the discipline itself. The GitHub and Gitea chapters underneath it are the place that layer lives: the repository is what makes it reviewable, shareable and undoable, which is the whole reason it is checked in rather than pasted.