I have to tell it a whole load of do and do not's after or before every prompt almost and even then it still does dumb shit - and this is despite the memory thing supposedly learning how to work with me. Still fun though, even if I got pissed at it for wasting my limited usage.
Discussion
yeah, the base $20/month subscription is pathetic. but the $80 pro sub is giving me more service than $200 got me with cursor.
i guarantee you, if i can't use up their pro subscription, neither can you. i've never spent so many hours working on code since 2019, when i was taking amphetamine and smoking weed all day long building an upgrade for a bitcoin fork called parallelcoin.
what i will say though is it helps a lot to have a good grounding in the things you expect claude to do for you, so you can correct it, and it takes some experience to learn what the typical fuckups are for this model. cursor's composer model is probably about on par but a lot more expensive. maybe composer does a tiny bit less dumbshit. probably it was trained partly by using claude as a trainer.
also, a really good thing about anthropic's offering is they partition your use into various time windows, and if you manage to use it, you can optionally pay for more but usually it will reset the session counter soon and be able to go again.
you literally would have to be asking it to do two things at once for days on end before you even get close to using up a session block.
I limit my user and try to spread it out but hit the weekly limit by Monday morning and had to wait till Tuesday eve for the weekly reset. I was very tempted to pay for extra use but it's difficult to judge what you'd get for your money like that.