[nextaccountic]

I have this superstition that if I praise correct outputs, the model may use this information that previous output was correct to produce better answers in future turns

But this only makes sense in the context of a new ask, and not as a standalone message