I am having better luck having it output blocks of new code and then summarizing it as a diff. I think this is another case where the model has processing gain

I am having better luck having it output blocks of new code and then summarizing it as a diff. I think this is another case where the model has processing gain when it "thinks it through step by step" versus trying to jump to the final answer in the form of a diff.