Nowadays using AI when writing software is sometimes more of a requirement instead of an option. Building without it, it may be impossible to deliver projects at the same speed as other teams. However, I think one of the biggest concerns is that using coding tools like Claude Code, Codex, Cursor, etc. really calcifies one's ability to code by hand and to solve problems.

Review the code that the AI wrote

It is greatly important to actually know and own the codebase. This is getting really hard when vibing. Sometimes it gets to a point that nobody knows except Claude what this code does. It can lead to a dangerous situation that when the AI server goes down, the project development grinds to a halt or that even the AI doesn't even know what spaghetti it wrote and then nobody knows. I like to treat coding assistants similarly to a different developer submitting a pull request. This way I still somewhat know what is going on in the codebase and it's easy to contribute to it myself without AI

Use AI as a last resort on hobby projects

If you find yourself using AI a lot at work, try to use it a bit less on personal projects. The biggest dopamine hit when programming for me always was that Aha! moment after solving a difficult problem. It kept making projects and games interesting, but when using AI to solve a bug that feeling is largely gone. Solving those bugs manually also gives the coder more domain level knowledge, it deepens their understanding of how the programming language actually works, how the system operates, software architecture, etc.

Well, but not using AI completely in this day and age is also stupid. It's an amazing technology, although it's also dangerous. I've made a rule for hobby coding: only use AI if I can't solve a bug after 30 minutes. I don't want to spend my whole day debugging a simple problem, like it was common 5 years ago. This way I still sometimes get that rewarding feeling of solving problems if I manage to solve a problem on time, but don't spend an entire afternoon searching StackOverflow and the reference

Read the docs instead of asking ChatGPT

I know it's very tempting, but I actually prefer to read the docs instead of asking ChatGPT some language concepts that I'm not familiar with. The programming language reference tells you everything you need to know. LLMs can still sometimes make up stuff. One time I was asking how to deploy something using Azure CLI and it gave me a command that didn't even exist! It may take a little bit longer to find the description of the functionality in the docs than asking an AI, but I think it's worth those 10 extra seconds.

So in summary, understand what the AI is writing, know your stuff in order to keep the code quality and actually use LLMs to their full potential.