It’s in my cursor rules. I sometimes even directly tell it in my requests. No matter what, it still says it. This is everything that’s wrong with consumer AI.
It’s not cursor. It’s the model. You can add context that tailors its output, but it doesn’t always fully respect it.
I have one model say, “Right on!” instead 
3 Likes
My favorite is:
“Now I can see the issue!”
…
[after 20 seconds of more thinking after it realizes it clearly didn’t see the issue]
…
“Now I see the issue!”
…
[more thinking]
…
“Now I understand the issue much better!”
…
[more thinking]
…
“Perfect! […]”
…
[fails to implement the wrong solution it came up with]
…
“Let me try a different approach.”
:'-D
3 Likes
that’s just its internal prompt set showing you it stoachstically in one response is actually a series of inferences with some compounding random error. good prompting makes this really useful, but we should really know when it is happening because it has implications for how we should trust responses and treat them temporally given our understanding of human communication.
Don’t forget, “Ah, I see the problem now. Thanks!” Claude once created six versions of the same file, with names like “-optimized-final.go,” “-efficient.go,” etc., because it got fed up trying to fix its own code that it generated
in most such cases its best to go to previous request above where the issue occured and adjust the request, then let the AI do the improved implementation instead of just continuing with increased context and hallucinations.
Most likely a fresh chat helps as well.
Some of the language used insinuates it has a backend chain of thought they aren’t showing, they don’t include it I assume due to model verbosity recently, but in general, it’s not out of nowhere. There is a model orchestrator or obviously a very large chain of logic, programming, or prompting we don’t see.
My recent favorite was Claude 4 exclaiming, “EUREKA! I found the smoking gun!” while searching for a bug. [Editor’s note: Claude 4 had not found the bug.]
1 Like