A Year With AI Coding Assistants: The Quiet Wins
After a year of pair-programming with AI, the biggest benefits weren't the ones the demos show. They were the boring ones.
I've used an AI coding assistant almost every workday for a year. If you'd asked me in month one what it was good for, I'd have said autocomplete on steroids. A year later, that's the least interesting part.
The win I didn't expect: explaining code to the model forces me to slow down. When a function won't explain cleanly, it usually means the function is a mess. I've caught more design problems by describing them than by reviewing.
It's also a patient rubber duck. I'll paste a stack trace and half a rant, and the act of writing the question often shows me the bug before I finish. The answer is secondary.
Where it struggles hasn't changed much. Anything that needs to know the whole system, or the reason behind a choice, gets shaky. It will confidently edit the wrong file if you let it.
My rule now: let it write the small, well-bounded stuff, tests, boilerplate, a one-off script. Review everything that touches data or money. The speedup is real, but trust is earned line by line.
One more thing. The assistants made me a slightly lazier reader of docs. I'm working on that, because the model is only as current as its training, and last week's API change is exactly the thing it gets wrong.