While I agree with the outcome, I don't agree with reasoning. What really matters for larger projects isn't how quickly you can write code, but how easy it is to understand the code and the architecture to those modifying the specific part. And this is where exclusively relying on LLMs is a disservice to you in the end — while they are good at generating _plausibly looking_ code, it's typically quite flawed, but in a non-obvious way, which is often hard to catch even for senior engineers. Thus for longer term it's actually beneficial to write most of the critical code by hand, and only leave less important stuff to LLMs. What percentage should be left to LLMs highly varies between projects, so there is no single good number.