On: Chain-of-thought can hurt performance on tasks where thinking makes humans worse

[fxnn]

Maybe it needs even smaller steps, and a programmatic (i.e. multi prompt) habit to always double-check / validate the assumptions and outcomes.