I think good results come from spending a substantial amount of time reading the claude-code documentation and setting up an environment that lets you work in this new paradigm. Asking it a few things, getting poor results, and deciding "it doesn't work when you ask it to correct itself won't get you that far.