Your idea of using a hillclimbing loop for improving documents is interesting. It looks like you found problems by looking at conversation records. Can you explain the measurements you use to know if a version of a document solves a problem? Also what do you do when different coding assistants give comments, about the same document?