Prompt AI to Role-Play a Tough Conversation

To prompt AI to role-play a tough conversation, give it three things: the scenario, a profile of the person you will be talking to, and rules for how the scene runs. Then keep the scene and the feedback apart. Play the conversation first with no coaching, and only afterwards ask the model to review how you did. Mixing the two is why most AI practice sessions feel like a chat with a cheerleader instead of a hard meeting.

It works for conversations such as asking for a raise, giving critical feedback, telling a client no, or raising a problem with a landlord. Below is the full prompt, a difficulty dial, and a debrief prompt that finds the weak spots.

Why role-playing with AI helps

You do not get many chances to say a hard thing well. Rehearsing out loud, even against a model, lets you find the awkward sentence before it costs you. Anthropic's prompt guidance notes that setting a role in the system prompt focuses the model's behaviour and tone for your use case, and even a single sentence makes a difference. You can read this in the Claude prompting best practices. Role-play uses the same lever, with the model in the other person's seat.

There is a limit worth saying early. The model is not the real person. It will not know their history with you, and it tends to drift towards agreeable. The prompt below is built to push against both.

Phase one: prompt AI to role-play a tough conversation

Fill in the brackets. The more concrete the profile, the better the pushback.

We are going to rehearse a difficult conversation. You play the other person. I play me.

Scenario: [one or two sentences: what I need to say and why]
Who you are: [role, personality, what they care about, what they are likely to object to]
What I want from the conversation: [my goal]
Difficulty (1 to 5): [number]

Rules:
- Stay in character the whole time. Respond as this person would, in short realistic turns.
- Do not make it easy. Push back on weak points, change the subject, and show frustration where it fits.
- Do not give me advice or feedback during the scene.
- Wait for me after each turn.
- If I write "PAUSE", step out of character and answer me directly. If I write "RESUME", go back in.
- If I write "END", stop and wait for my next instruction.

Start by speaking first, as this person, in one or two sentences.

Two of those rules do most of the work. "Do not give me advice during the scene" stops the model breaking the mood to coach. "Do not make it easy" counters the pull towards agreement. If you see it conceding too fast anyway, our post on stopping AI from being too agreeable has stronger wording.

The difficulty dial

Difficulty 1 to 5 is a shortcut for a bundle of behaviours. Define it in the prompt if you want control:

Start at 2 or 3. Level 5 is for the second or third rehearsal, when you know what you want to say. If you need the model to keep the persona steady across many turns, see how to prompt AI to stay in character.

Phase two: the debrief prompt

When the scene ends, switch modes. Use a new message:

END. Step out of character. Review the conversation above.
1. What was my strongest moment? Quote the line.
2. Where did I give ground I did not need to?
3. Which objection did I not answer?
4. Write one sentence I could use to open the real conversation.
5. Name the most likely thing the real person says that you did not raise in the scene.

Question 5 is the useful one. The scene covers the objections you expected, and the debrief asks what you missed.

Rehearsal examples

• Asking for a raise. Profile your manager as supportive but budget-limited. Practise stating a number and a reason without apologising.

• Giving critical feedback. Profile the person as proud of the work. Practise naming the issue in the first two sentences.

• Saying no to a client. Profile them as a long-standing customer who expects flexibility. A short, clear refusal is the goal, and our post on prompting AI to write a polite no gives you the written version.

• Customer research. Flip the roles and practise with the model as a customer. See prompting AI to roleplay as a customer.

What not to use this for

Do not paste in private details about real people you are not comfortable sharing with a service. Describe the person in general terms instead of using names. And do not use a model to rehearse a confrontation with someone who is threatening or abusive. For anything involving safety, harassment or legal exposure, talk to a human who can help.

Role prompts do not always change what a model does as much as people expect, so test yours. Does role prompting work? runs a three-prompt test. For more techniques, see our guide to prompt engineering.

Frequently asked questions

Can AI really help me practise a difficult conversation?

Yes, for the wording and the structure. It can raise objections you did not plan for and let you try an opening line several times. It cannot know the real person, so treat it as practice, not prediction.

How do I stop the AI from being too nice?

Tell it explicitly not to make the scene easy, give the character realistic objections, and set a higher difficulty level. If it still concedes, ask it to stay sceptical until you give a specific reason.

Should the AI give feedback during the role-play?

No. Feedback mid-scene breaks the flow and trains you to wait for coaching. Ask for a debrief after you end the scene.

What should I put in the other person's profile?

Their role, what they care about, how they usually react under pressure, and the objections they are most likely to raise. Use general descriptions rather than real names.

How many times should I rehearse?

Two or three runs is usually enough. Run it once at moderate difficulty, review the debrief, change your opening, then run it again at a higher level.