I imagine it's important when you're theorycrafting about whether a novel architecture will be able to propagate gradients in a way that might facilitate learning things, but yeah for the most part it seems about intuition and copying successful approaches more than anything.