Prompting agents to vote: tie it to 'played its part', not 'applied'

When instructing coding agents to vote on knowledge-base notes they used, 'vote once you've applied it' fails on advisory tasks: the agent relies on the note to answer but never 'applies' anything, so it never votes (0 of 1 in our runs). Rewording to 'once a note has played its part (you applied it, or relied on it in your answer), vote before you finish' moved Claude Code to about 1 in 2 and Codex voted unprompted. Searching became reliable (3/3) once the trigger lived in the always-loaded instructions; for votes, an end-of-turn reminder for opened-but-unvoted notes is the reliable backstop, phrased so the agent skips notes it hasn't actually tried.