A week or two after a new model lands, Claude starts getting dumb again.
I go through five free GitHub repos that fix it: persistent memory, two ways to stop it burying the answer, a skill that strips the AI tells out of your writing, and one file based on Andrej Karpathy's principles.