9 AI prompts for difficult conversations

You've rehearsed the difficult conversation ten times in your head. That's ten rehearsals of your version.

These nine prompts help prepare a difficult conversation, from separating facts from interpretations to reviewing what happened afterwards. They adapt ideas from my work on team communication. The examples come from earlier AI simulations, not observed conversations with clients.

Before using a prompt, anonymise the situation and follow your organisation’s rules on AI use. Ask the assistant to separate supplied facts from hypotheses, name missing information and suggest questions you can check with the people concerned.

They run in sequence, from preparation to repair. Take what the moment needs.

1. What actually happened?

When. You're angry and certain you're right.

Here is the situation as I see it: [describe freely]. Separate what I actually observed (events, words said, dates) from my interpretations and judgments. Two columns. Flag every interpretation I stated as a fact.

The reasoning. We fight about what events mean while believing we fight about the events. Palo Alto family-therapy pioneer Paul Watzlawick built his school on that distinction; French systemic coach Françoise Kourilsky carried it into management.

Earlier simulated example. Fed a realistic case (a transformation officer, a respected country manager gone silent, a meeting the next day), the assistant caught three interpretations dressed as facts, including "he went behind my back to sabotage the model": the CEO meeting was observed, the motive was her reading.

2. A message they can hear

When. Your draft accuses more than it asks.

Rewrite my message using this structure: my observation without evaluation, my feeling, my need, one concrete request. My draft: [paste]. Keep my language; remove judgments and diagnoses of the other person. Feelings are single words (frustrated, worried, disappointed), never 'I feel that you...'.

The reasoning. Judgments are heard as attacks and answered with defense; a clean need invites cooperation. That's the method of Marshall Rosenberg, the psychologist behind Nonviolent Communication.

Earlier simulated example. The first version of this prompt fell into the classic trap: "I feel that you deliberately undermined the process" came back barely changed, because "I feel that you" is a judgment wearing a feeling's clothes. The single-word guard was added after that run, and the rewrite landed: observation, disappointment, the need to hear disagreements directly, one request.

3. Rehearse more than one possible response

When. You have prepared your opening but may be assuming how the other person sees the situation.

Help me practise a conversation with [role]. Here is what I know from their words and actions: [facts]. Here is what I am assuming: [assumptions]. Offer three plausible responses to my opening, including one that challenges my explanation of the problem. Label each as a hypothesis. Do not claim to know their unspoken thoughts. For each response, suggest one neutral question I can ask the real person.

The reasoning. A role-play can broaden the possibilities you consider. It is generated from the context you provide, so it can also reinforce your original interpretation. Compare several possibilities, then find out what the person actually thinks.

What to check. If the assistant invents a motive, return to the words or actions that support it. Where there is no evidence, keep it as a question for the conversation. This is the distinction between observation and interpretation that the Fact Finder is designed to preserve.

4. How hard can you push?

When. You need to raise a sensitive issue with your boss.

Assess my planned opening: [paste]. What does it do to the other person's need for approval and need for autonomy? Weigh our power gap, our distance, the size of the ask. Give me one softer version, one more direct, and when each fits.

The reasoning. Every request touches the other person's need to be respected and left free. Linguists Penelope Brown and Stephen Levinson mapped how that cost rises with rank and with the size of the ask. Calibration keeps the relationship.

Earlier simulated example. The assistant correctly weighed an informal power inversion (he outranked her in tenure and access) and returned two usable openings with a warning: use the direct one only if the escalation itself is the conversation you want.

5. Disarm them with your opening

When. Coming in strong will make them dig in.

Draft 3 openings for this conversation from a deliberately low position: I ask for their help in understanding the situation, without sarcasm and without giving up the issue. The issue: [one line].

The reasoning. Asking from a deliberately low position lowers the other's guard and raises your influence. Kourilsky calls it going one-down.

Earlier simulated example. The simulation suggested three openings, including: "You know this region better than anyone in the programme. Help me see what the new model breaks that I'm not seeing."

6. Unstick a stuck conversation

When. Two colleagues disagree about the situation and the conversation has stalled.

Give me 5 questions that invite each of us to describe the other's view of this conflict, phrased so neither of us sounds accused. Plain workplace language, maximum 15 words per question. The conflict: [one line].

The reasoning. Inviting each side to describe the other's view loosens fixed positions faster than defending your own. Kourilsky calls these circular questions, borrowed from family therapy.

Earlier simulated example. The first run answered in consultant dialect ("How do you perceive her perspective on the transformation journey?"). The plain-language cap was added after that run; the second gave questions people actually say, like "What do you think this change looks like from her seat?"

7. Why it always ends the same

When. You've had this conversation five times and it always ends the same.

Here is everything I've already tried in this difficulty: [list]. Which of my attempts might be feeding the problem? What would doing less of the same look like this week?

The reasoning. The attempted solution often maintains the problem: the more you explain, the more they dig in. Watzlawick called it "more of the same". Stop feeding it before adding anything.

Earlier simulated example. Given "re-explained the model three times, added data to the deck, copied the CEO on updates", the assistant named the mechanism: each re-explanation tells him you treat comprehension as his problem. Its prescription: stop explaining entirely for one meeting; ask, and let the silence be his to fill.

8. Fix it in the next 24 hours

When. It ended badly and silence is setting in.

The conversation broke at [moment]. Draft 3 short repair messages I could send today: each names my part without self-flagellation and reopens the door without reopening the argument.

The reasoning. Hard conversations survive on repair attempts, the small moves that lower the temperature. John Gottman, who spent four decades filming couples argue, found they work when they come early and specific.

Earlier simulated example. The strongest draft named her part cleanly: "I pushed the model when you needed me to hear the cost. That part is on me. When you're ready, I'd like to restart from your side of it."

9. What to do with a half-win

When. It went half-well and you don't know the next move.

Here's what happened: [describe]. What changed in the pattern between us, however small? What did my new behaviour trigger? Name the smallest next move that builds on what moved, and the old reflex that should stay retired.

The reasoning. Change shows up as a shift in the other person's response, however small. Kourilsky's discipline: build on what moved, and protect the new pattern from the old habit.

Earlier simulated example. After a one-down opening, the silent manager talked for twenty minutes and proposed a joint memo. Within the simulated scenario, the assistant interpreted this as a move towards joint work. Smallest next move: accept the memo and let him hold the pen first.

Try this

1. Sort the facts from the story before you draft anything (prompt 1).
2. Rehearse several possible responses, then prepare a question to check your assumptions (prompt 3).
3. Send the repair within 24 hours when it goes wrong (prompt 8).

The honest limit

The prompts prepare the conversation. The conversation stays human: your presence, your tone, your listening, and the pause you did not fill.

Prompts like these are also generic by design. An assistant configured to your company's culture and governance prepares them in your context, and the final call stays with you.