I think this might work for smaller codebases, but the main point of my article isn't really about vibe coding. Vibe coded apps are typically smaller anyway, so refactoring isn't that big of an issue there.

When we're talking about actual software that has been around for a while and has accumulated serious tech debt, it's not so easy. I've definitely worked on apps where your described approach doesn't lead to anything viable. It's just too much for an AI to grasp when you have years of accumulated complexity, dependencies, and business logic spread across a large codebase.

Regarding vibe coders specifically: I think people who can't code themselves often don't really know what "cleaner design" or "more reuse" actually means in practice. They can certainly learn, but once they do, they're probably not vibe coders anymore.

With AI generation of code or text, I have found that quality improvements have to be run multiple times, successively, until the achieved quality reaches my expectations. Prompts must also be refined before letting it run multiple times.

In my experience codex fails at that kind of task in many cases.

I think it’s unlikely it will succeed at that task with a real world , old, large codebase.

Now, the next release might work !

And who's going to tell the managers?