Last week or so we got the message at the end of the day that all claude models were performing poorly, this was after I worked on claude models the whole day swearing because of performing less then a 2 year old, I call it Embryo brain performance.
This is again happening today…
No matter how I describe my prompts it simply ignores most of it, rules are ignored etc. I have to explain 10 times the same prompt and it still makes excuses that it didnt follow the other 9 times. This is so frustrating…
I have recently experience Sonnet starting to ignore rules and even ignore TODOs in a plan (specifically ones where the agent is supposed to pause all work and await the user to instruct when to proceed).
The latter doesn’t bother me much because it probably depends on the wording of those TODOs but I haven’t studied hard enough yet to figure out what wording helps make it absolute.
But Sonnet ignoring rules seems more prevalent right now. And it’s definitely not because the rules aren’t in the context – I’ve checked…and have even had discussions about it with Sonnet, asking why it’s skipping rules. It ultimately says “oops, my bad, I’ll do better next time” . But we know how stupid that is – these LLMs think they are immortal and will be able to do better with a fresh clean context – LOL.
Oh, and this was definitely with Sonnet 5. It happened numerous times with me yesterday. (I wasn’t using Sonnet 4.x since there is the Sonnet 5 discount going on right now.)
Today I’ve actually just been using Composer 2.5 for everything because I’ve been trying to convince myself to start using it more and give it a chance. So far today, it’s been flawless – and I never give these things easy things to work on.