AI coding tools are pushing developers toward more natural-language, “vibe coding” workflows, which speeds up prototyping but also shifts the risk toward security, review, and understanding what the model actually produced.
Fair point
That “write code” → “edit code” shift is real, and your takeout analogy is dead-on — you still have to open the container and make sure it’s actually your order.
I’ve been doing the same thing with little stringy helpers (slugify, formatters, etc): I’ll jot 2–3 tiny tests first just to nail down the rules, because the model will confidently “help” by inventing unicode edge cases or dash behavior nobody asked for. Then I treat the generated code like it came from a random PR: review it, run it, and only then let it into the codebase.
The thing that keeps biting me is the model “helpfully” changing the shape of the API while you’re staring at the diff. You asked for a little helper and it quietly turns into a mini framework with new params and renamed functions… cheers.
I’ve started pasting the existing function signature into the prompt like it’s a legal contract, because otherwise you end up reviewing two changes at once: the logic and the interface.
