Claude Code is an Agentic Loop agent. It will figure things out on its own given the task and context. We should avoid giving it concrete instructions of how to do things. Our focus should be expressing our intent as clear as possible to Claude Code, giving it enough context. There is no hard rule of how much is enough. It is context dependent.

Started with Claude Code last week on a large project, with more than 7 years old, 4 years in production. As recommended from Anthropic, I started with "/init" command. A CLAUDE.md file content was proposed. The content looked promising. Claude Code picked up some notions from the code and assumed that the project was "Layered Architecture + CQRS + Clean Architecture"; and the business logic stayed in "Aggregate" at the Core layer. It is legitimate to assume that. Developers who read for terms and patterns will assume the same.

It is not accurate. Hence, I told it:

The project follows a Slice (Feature) Architecture. It is organized in Layers. It uses the notion of Command and Query. However, it is not CQRS pattern that we are described in the internet. All the business logic and validation are handled at the Application layer. The Aggregate ensures the state consistency because we store states with Event Sourcing instead of relational database. ...
It is NOT a Clean Architecture.

I got a useful CLAUDE.md file.

I have seen developers used Claude Code with predefined Skills and other markdown files. They are trying to give Claude Code more context and instructions. I am very much in doubt its usefullness. There are a few concerns. How do we know they are accurate? How to keep them up to date? Is it true that more information is better? The more information we feed AI (think of human here as well), the more likely they are conflict or cause confusion. We have to manage that. Keeping things in sync is not an easy task.

My line of thought and approach is simple. I have Claude Code with a master CLAUDE.md file. Source code is the single source of truth. Me is the context provider and the master who will ask questions. I do not have to worry about Claude Code. I do not have to worry about source code. It is always there with the branch I want to work on. The only concern is ME asking questions. I need to learn and practice the art of asking questions. AI does not get offended or mad. So, emotion is out.

Practice the art of asking questions.

It is an amazing tool. I have completed so many things with it. I have learned a ton with it along the way. I even built a CLI application that acts as my personal training coach with Claude Code, written in Python. I do not know Python. The app just works perfectly.

Well, that is enough. Until next time.