Ask HN: Is Claude Code less useful in recent weeks for you?
In the past couple of weeks, I've found Claude Code much harder to work with. It seems to not understand well-known (eg for me, .Net) libraries as well as it should, and I have had to download the library source from github and instruct it to read the library to know how to use it; it does not follow instructions as well; it frequently assures me something is fixed but sometimes has forgotten a key portion; and it frequently ignores its memory: for example, it commits even with a stored memory instruction to never commit without me verifying it first. It's starting to be a real time drag and I am not getting the productivity I did even a short time ago. Is anyone else seeing this?
12 comments
There's a lot of people complaining about degradation for weeks on /r/claude.
I do wonder if it's more that it's just lost the wow factor and they're seeing the cracks as I generally still feel it's at the same capability as I'm used to. But I've always viewed LLM code with skepticism and rewrite a lot of it.
As someone who's mainly .Net/Typescript I'm not seeing it be terrible with libraries, but it's generally bad when libraries constantly change their API. React Router for example. So depends what libraries you're using. We use MaterialUI and trying to get it to consitently use the new Grid definition is an exercise in futility.
Do you know about the context 7 MCP, that can help apparently though I've not really bothered with it myself. Not sure how many .Net libraries they have.
I do wonder if it's more that it's just lost the wow factor and they're seeing the cracks as I generally still feel it's at the same capability as I'm used to. But I've always viewed LLM code with skepticism and rewrite a lot of it.
As someone who's mainly .Net/Typescript I'm not seeing it be terrible with libraries, but it's generally bad when libraries constantly change their API. React Router for example. So depends what libraries you're using. We use MaterialUI and trying to get it to consitently use the new Grid definition is an exercise in futility.
Do you know about the context 7 MCP, that can help apparently though I've not really bothered with it myself. Not sure how many .Net libraries they have.
Great tip, I'll check that out. Thankyou.
Claude has suffered some major bugs. Here is Anthropic's post-mortem.
https://www.anthropic.com/engineering/a-postmortem-of-three-...
https://www.anthropic.com/engineering/a-postmortem-of-three-...
It’s gotten noticeably worse on my projects as well. Likely going to switch to Codex until they also start degrading performance for better margins
Unfortunately, for major use it is twice as expensive ($200 / month vs $100.)
It's been going up and down, with the general trend of going down, over recent weeks. Oddly, it also seems to have been changing it's coding style (at least with Python and TS) which I find somewhat humorous.
I chalked it up to Anthropic engineers tweaking stuff in the background a lot.
I chalked it up to Anthropic engineers tweaking stuff in the background a lot.
I haven't been fully keeping up, but anyone else notice the same with Cursor? Could be because I use Claude models (GPT-5 was too slow and iffy results). Funnily enough, Codex has been really impressive recently and I've swapped almost completely to using Codex.
I've noticed something similar, especially when dealing with specific bugs
Chatgpt is what I noticed has gotten significantly worse
enshittification is the only true constant under capitalism
While I agree with you (very visible in recent years) I'm not convinced -- I don't think Anthropic's moat is high and it would be a very risky situation this early in their product's life.
I have to admit I have wondered if they are cutting costs, though, with side effects. It's the natural suspicion :(
I have to admit I have wondered if they are cutting costs, though, with side effects. It's the natural suspicion :(