Every developer who uses AI has felt this.

You open a new chat.

You explain your architecture.

You explain your file naming.

You explain that you do not use repositories, prefer a rich domain model, and that your handlers and services return a Result<T> .

The AI writes good code.

The next day, you open another chat, and you explain it all over again.

This is the hidden tax of AI-assisted development.

You repeat the same instructions in every session for every feature.

A CLAUDE.md file helps; it's added to every session, but you can't make it with a lot of instructions (it should be 100 lines or less).

It holds the facts about your project and overall architecture, not the step-by-step recipe for every kind of task you do.

There is a better way: skills .

A skill packages a set of instructions, and even example files, that Claude loads only when it needs them.

You write the recipe once, and Claude follows it every time, in every session.

Skills also work if you are using the OpenAI Codex…