Does it help to curse at Claude Code, because oh my god does it make some questionable choices. So far I am holding my tongue because I've also heard that it can spiral.
"I want each sentence on a separate line."
*writes code to concatenate all sentences, which is then repeated on n lines where n is the number of sentences*
"No really I want each sentence on a line by itself"
*generates 5 pages deliberating on what I mean by that, considering the possibility that I am confused and don't actually want that at all, then going back to deliberating on what I mean it for 5 more pages*
Also my laptop crashed recently, and upon restore, I am unclear whether it remembers anything. Claude's personality also appears to be completely different. It used to speak in short sentences and a lot of bullet points, now it is writing very long paragraphs with no bullet points.
Don't get me wrong. I'm not trying to mock coding assistants. I'm using Claude Code to explore some initial hypotheses through visualization. I am not very sophisticated when it comes to creating visualizations, and I've done in a few days what would have taken me a very long time.
(it would take me a very long time because I don't have a lot of time to do coding myself)
I am able to explore these hypotheses before I send my research team down a rabbit hole. So that is good. Any my hypotheses are mostly panning out. Net positive, net win.
It comes at the cost of the most frustrating things that I feel like no student would ever misunderstand. And students don't "compress" every few hours and forget important stuff. But I can iterate quickly, so it is probably still a win.
I've gained a lot of clarity on what parts of computing science are likely to remain valuable for the foreseeable future. E.g. I helped Claude work through some issues it was having with computational complexity because it was making some poor choices. I knew what to ask and where to point it.