How I give coding tasks to AI

I tried AI autocomplete, but the constant suggestions often interrupted me. While I was figuring out what the code should do, I also had to decide whether to accept the next generated line.

I get more out of AI when I give it a small, complete task. I describe the change, its limits, and how I'll check it. A request might look like this: Update this file to use the new interface. Keep the existing behavior and make sure the tests still pass.

That gives me something specific to review. A request like "make this better" leaves too many decisions open, especially when I haven't decided what better means myself.

Once I have a draft, I read the diff and check that it matches the request. I look for effects outside the changed code. I run the relevant tests. I merge when I understand the change and the checks support it.

Compiling is useful evidence, but it doesn't tell me whether a change fits the codebase. That's still part of my review. If I can't explain what I'm asking AI to do, I need to spend more time understanding the problem first.