I was doing a personal hobby thing with Claude, because personal hobby projects can go from idea to fun thing to play with in a matter of hours now. Problem is, Claude has a tendency to make that… not quite an hours-long job.

Claude is said to have “good judgement” from many people, and I feel that way too. It’s not really measurable from benchmarks but it feels like it. But that sometimes backfires if I’m doing casual things - it spends a lot of time architecting assuming a data error is a disaster, one mistake will cost me dearly, and more… even just for a small personal thing I wouldn’t mind breaking occasionally.

It is useful when I’m trying not to, say, cause a security vulnerability or making something that will see public usage like the monitor or this blog itself - but it also gets a bit repetitive saying to Claude that “this is a hobby project, it shouldn’t matter too much”. Asking Claude to remember it in project memory seems to work pretty well though.

As a side note, I did notice Sonnet is a bit more likely to “yolo” things, or do things that Opus might’ve warned me many times over. Either a pro or a con depending on what you’re doing, and either way the auto mode classifiers usually block anything too risky. I might try some non-Claude models more soon though!