💭 if you can cheaply and effectively score responses does automatic prompt optimization pretty much solve the problem?