The AI (for instance any AI, but I have use Claude Sonnet 3.5 mostly), is losing context after 2 prompts. It used to be WAY BETTER.
I’ll give an example:
Claude 3.7 says: (cuz I need to reference a dylib I am creating in a diff location depending on the type of build):
-
Add the library search path: $(PROJECT_DIR)/Xbuild/$(CONFIGURATION)-$(PLATFORM_NAME)
-
Add to Other Linker Flags: -lSudokuSwiftBridge
I show it the error generated. It then reverts to FORGETTING COMPLETELY that I need to reference the dylib in diff locations.
And I say: “as you know the build location is different depending on type of build and phone vs simulator - you seem to forget that even though you know the script of course does that since you generated the script”
And it apologized - but it lost that context IMMEDIATELY. THIS IS NEW BEHAVIOR - I WANT TO EMPHASIZE THAT BECAUSE IT MATTERS A TON.
Then a couple of prompts on it tells me to go back to the bash shell script that generated the error. Again: IF IT HAD CONTEXT AT ALL IT WOULD NOT KEEP REVERTING TO PREVIOUSLY ERRONEOUS SOLUTIONS.
To which I say: “did you already forget the the sandbox is preventing us doing this? what happened to you?” and Claude 3.7 apologizes again.
WHAT HAPPENED? Things used to work great, and in 0.46 the context is gone - it’s just not here - I’m not gonna mince words here. IT IS COMPLETELY BROKEN.
It used to maintain context for 20 to 30 prompts - I use it intensively for the last 2 months to generate 300k of code - I know it’s broken cuz I remember how it used to actually work.
THIS IS REALLY DISAPPOINTING. FIX YOUR PRODUCT, PLEASE.