Yea, this is one thing that gets me. Vibe coding and freestyle prompting gets wildly different results. Starting from a set of specs and expectations routinely gets me reviewable and we'll tested code. Sure, running the full test suite alot slows things down and uses more tokens, but when I hit a five hour or weekly window, I just roll up my sleeves and start working without the model because the code is layed out the way I expected and is designed to fit my meatbag workflow and pace.

You must not be looking very closely yourself then.