My current favorite practice is do only the most important thing now. That's usually "make it work", not "make it fast". It's much easier, for me, to optimize a program that produces the correct output than to fix a very fast program that produces incorrect output.