I Read the Generated Code Twice as Slowly

When the model started writing whole functions for me, I noticed a strange reflex: I trusted its code more than my own. It came out polished, confident, instantly. Surely something that fluent was right.

That reflex cost me. The code looked correct and read beautifully and quietly assumed a thing that wasn’t true. I’d approved it in seconds because I hadn’t struggled for it — and the struggle is exactly what usually builds the model in my head that catches the mistake.

So I made a rule for myself. Code I didn’t write, I read more slowly, not less. Like a stranger sent it and I don’t quite trust them yet. Where does this fail? What did it assume about my data? Which edge case did it skip because I never mentioned it?

The machine made writing cheap. It did nothing to make reading cheap, and reading is where correctness lives.

I still let it type. I just never let it decide. The generated code is only ever as good as the person willing to slow down and actually read it.

– Serguey Asael Shinder

Leave a comment