Using Claude as a second brain while coding

I've used every generation of AI coding tools - Copilot when it launched, various chat models, now Claude. My workflow has changed more in the last year than in the previous five. But not in the way most people describe it.

The common framing is autocomplete at scale: write faster, get boilerplate out of the way, don't look up API docs. That's real. But it's the least interesting part.

The actual shift

What changed how I think is using AI as a second pass on my own reasoning. Not to write the code - to interrogate whether I'm solving the right problem. I'll describe a system I'm designing and ask what I'm missing. I'll explain my data model and ask what breaks at scale. I'll paste a function and ask why someone would regret it in six months.

It's not that the answers are always right. It's that being forced to explain the problem clearly enough for a model to engage with it often reveals the flaw before the answer arrives.

What I've learned to avoid

Using it to generate code I don't understand. That's the trap. You ship faster, you understand less, you're slower the next time something breaks. The leverage is in amplifying your own understanding, not replacing it.

I still write the first draft by hand, usually. Then I use AI to pressure-test it. That ordering matters.