Most developers use AI coding tools wrong. I was no exception.
I tried AI assistants, got frustrated when they produced buggy code, and swore them off. Then I stumbled on Karpathy's framework for thinking about AI as a coding partner, and everything shifted.
The four principles that changed my workflow:
-
Use AI to draft and accelerate coding, but keep human responsibility for correctness. Treat AI as a first-draft generator, not a final answer. It writes code fast; you verify faster.
-
Delegate boilerplate, repetitive edits, and test scaffolding to the model. Use it for the boring stuff (boilerplate, refactoring, test generation) so you can focus on the hard problems.
-
Always review what it produces. Every single line. This isn't optional; it's the harness that makes the tool safe. Review AI output carefully for important or persistent code, even if Karpathy's original vibe-coding demo allowed "Accept All" for disposable projects.
-
Iterate in small loops. Prompt, review, correct, repeat. Don't dump a whole project and hope.
Once I applied these, I couldn't go back. The speed difference is real, but only because I stopped treating AI as a replacement for my judgment and started treating it as a lever for my attention.
What's your experience? Are you fighting AI or working with it?