Notes on coding with AI

I’ve been writing code with AI assistance daily for a while now. A few observations that have held up.

What changes

The cost of trying things drops to near zero. I used to weigh whether a refactor was worth attempting. Now I just try it — the model writes a draft in a minute and I can read the diff before deciding. The cost of the exploration changed, even when the final code is mine.

The bottleneck moves to taste and review. Anyone who can read code carefully and articulate what they want is suddenly more productive. Anyone who can’t is producing more code, faster, with no idea whether it’s right.

What doesn’t change

Understanding the system still matters. The model can write a function. It cannot tell you whether that function should exist, or whether it fits the architecture you’re going to want in six months. That’s still on you.

Reading is harder than writing. This was already true. AI just shifts more work onto reading.

// Easy to write, easy to miss in review:
items.forEach((item) => {
  process(item);
});
 
// Was the original loop sequential for a reason?
// Did `process` have side effects we cared about ordering?

Where I’m still figuring it out

When to not use it. There’s a class of problems where I learn by struggling, and the AI shortcuts the struggle. I haven’t found a clean rule yet. Right now I just notice when I’m typing prompts to avoid thinking, and try to stop.