Rehearse difficult conversations with an AI role-play and a feedback pass

Set up the role-play

Give it enough to be realistic, and nothing private it doesn't need:

I want to practice a hard conversation. You are playing
[who: a customer of eight years / a team member / a client].
What I know about them: [two or three lines].
What I need to say: [the point, in one sentence].
Push back the way they realistically would. Keep your
replies short, like real talk. Stay in character until
I type STOP.

Round one: just get through it

Say your part. Don't aim for perfect. Notice where you stall, apologize, or start explaining your whole business.

Round two: lead with the point

Start a fresh try. This time, say the main point in your first two sentences. Then listen to the pushback and answer it once, without circling back.

Step out of the role for feedback

Type STOP, then:

Step out of character. Tell me:
1. What did I avoid saying?
2. Where did I over-explain?
3. What is one sentence I should open with?
4. What is the hardest pushback I should be ready for?

Write down the opening sentence. That's the one thing to bring into the real conversation.

Do it out loud, not just typed

Typing is practice for writing. If the real conversation is spoken, practice by speaking. You can talk to ChatGPT like a phone call and run the same role-play with your voice.

Keep it fair

• The AI doesn't know the real person. It's guessing at how they'll react. The real person may surprise you, in either direction.
• Don't paste someone's private records, like a performance file or medical notes, into the chat. A few plain lines are enough.
• If the conversation involves letting someone go, a complaint about harassment, or anything with legal risk, follow your company's HR process or talk to an employment lawyer first. Rehearsal helps with your words, not with the rules.

Under the hood

Role-play works because models are good at producing plausible conversational responses to a described persona. The limits come from the same place:

• Persona fidelity is a guess. The model simulates a generic person matching your description, not the real one.
• Agreeableness leaks in. Models drift toward conceding. "Push back the way they realistically would" and "stay in character" counter that drift.
• Critique needs a mode switch. An explicit STOP plus a numbered critique prompt separates in-role replies from evaluation, so feedback isn't softened by the character.

Keep inputs minimal. A two-line description gets realistic pushback without moving personal records into a chat log.

Short close

Set up the role-play, rehearse twice out loud, ask what you avoided, and walk in with your first sentence ready.