Claude loves to claim it's done... sometimes it even is! The rest of the time, it is either close or does something you didn't want at all. The best way to find that out is by looking at the diff to see what it actually changed. If you want to let Claude Code act more autonomously, you can use the command /goal. You type a condition, and Claude will keep churning until that condition is true.
So what decides? Much like how auto-mode uses a small model to determine if an action is safe, /goal uses another model (Haiku, by default) to read the conversation and judge if the original goal was met, not met, or impossible. If it was not met, then it pushes Claude to continue until it is.
Since this "evaluator" can only see the conversation, it needs a concrete goal (much like humans)! Something like "Clean up the auth module" isn't very concrete, but "all tests in test/auth pass and npm run lint exits 0" is very specific and achievable.
Remember that using a goal doesn't change your permission mode, so if you are in Manual mode, Claude will still ask you for permission on each tool call; I still recommend using auto mode (Clue #19) most of the time!
Roscoe emailed the tip, Claude wrote the drafts, and I was the evaluator. The first two came back not yet met.