How to write prompts for code so that AI does not “break the project”?

You ask ChatGPT to write an authorization function. You get a beautiful code. You insert it into the project — and everything breaks. Sound familiar? The problem isn't the AI. The problem is how you talk to it.