Prompt design for AI personas that push back instead of agree

Most LLM apps optimize for helpfulness. Ours does the opposite. I'm building cosskill, a conversation practice tool. Users rehearse salary talks, breakups, and tough feedback before the real thing. A