I’ve found the best workflow has been to ask Claude to profile everything and identify the problem spots, then I have to be the one to propose the correct architecture, then Claude implements it. If I let Claude run wild it will invent all kinds of weird, weird hacks that compound the complexity.