I particularly like this framework: how hard is it to describe the task vs. how hard is it to check the output.