a prompt is a vote, not a checklist

A system prompt isn’t a list of rules the model dutifully ticks off. It’s closer to a vote among competing instructions — and a single soft mention loses to the paragraph leaning the other way.

I hit this translating formal Japanese to English. The prompt pushed hard toward natural, casual English, with one buried aside: “…unless the Japanese is clearly formal.” For keigo input, the casual pull out-voted that lone clause every time — deferential business Japanese came out as breezy texting English. The instruction was technically present; it just lost the vote.

The fix wasn’t adding more caveats. It was promoting the behaviour to a first-class rule with concrete recognition cues — naming the actual keigo markers (です／ます, 恐縮ですが, ～いただけますでしょうか) so the model couldn’t miss them. When a behaviour matters, make it load-bearing, not an “…unless…” tucked into a sentence pushing the other way.