I agree! I get much better results when I give LLMs some light architectural constraints such as to use tailwind and alpine.