Surprisingly, I don't use AI suggestions anymore. I find them distracting. I'm not sure when they started becoming an issue for me, but I toggled AI features off in my editors.
Instead, I prefer to use an agent in the terminal to do rough sketching for a feature, and fill the gaps myself in a distraction-free environment. I might even ask for validation ("be critical, not a sycophant") for an approach, without requesting any code changes. For people like me who work in isolation and build features e2e, having an assistant that operates this way is very helpful.
Of course, I'll delegate some chore work, but in principle I want to write my own code.
I find that this works best for me, because it avoids the three deadly AI sins:
- Making me lazy. I take pride in my work and don't want to push sloppy code.
- Causing my skills to atrophy. I don't want my brain to switch off. The model might have good ideas, but most of the time I have better ones. These come from real-world experience, and that's what differentiates me from some other AI operator.
- Making me lose the feeling of ownership. This is where teams fail. No one wants to take ownership, stuff falls through cracks, and bad things happen. In contrast, when I write the code, I know exactly why I did it this way, and it's my responsibility to maintain it.
So this is where I draw the line with AI in my workflow. Things might change, but I get a significant productivity boost without noticeable drawbacks.