> and hate how much extra bullshit I didn't ask for they always add to the output.

I can recommend for that problem to make the "jumps" smaller, e.g. "Add a react component for the profile section, just put a placeholder for now" instead of "add a user profile".

With coding LLMs there's a bit of a hidden "zoom" functionality by doing that, which can help calibrating the speed/involvment/thinking you and the LLM does.