“I could have said that more clearly.” This thought often only comes after a conversation. You can practise a form of words beforehand without the AI having to predict the real person you will be talking to. The goal is smaller and more tangible: stating your concern clearly, recognising an assumption and summarising an agreement cleanly.
ChatGPT Voice enables spoken conversations in supported interfaces. Microphone permission, account, rollout and workspace settings affect access. The following sequence is an exercise of our own with invented roles. No voice session was carried out for it.
Choose a situation with exactly one point of uncertainty
Our example is the preparation of an information stand. You have agreed to put the content together. Your fictional colleague additionally assumes that you will take care of transport. The two roles therefore have understandable but different expectations.
This situation is more helpful than the instruction “Practise difficult conversations”. You know what it is about and how progress would show: by the end, it should be clear whether transport had already been agreed or still needs to be decided. Real names and confidential details of the conflict are not needed for this.
Give the role-play a contract
Take on the invented role of a fellow club member.
We are talking about preparing an information stand.
You assume that I will also take care of transport.
I want to clarify my confirmed responsibility in a factual way.
Respond to what I say without scripting my answers.
At the end, I will summarise the agreement myself.
Assess specific wordings, not my personality.The brief does not prevent every unsuitable answer. But it makes clear which exercise you intend. A role-play that immediately reads out all the right sentences for you would give you little opportunity to try out an opening or a follow-up question yourself.
Choose three observable points: was the concern stated? Was there a genuine follow-up question? Was the agreement repeated at the end? These criteria relate to individual contributions to the conversation. They do not need any supposed measurement of personality or feelings.
Keep working on one form of words
One possible opening of your own is: “I’d like to clarify which part of the preparation is definitely my responsibility.” If the role takes transport for granted, you can ask: “Had that already been decided, or is it an assumption so far?” That makes the different states of knowledge the subject of the conversation.
Pause when a section becomes too long or unclear. Repeat the same moment with just one change, such as a shorter opening. A comparison between two concrete wordings is easier to judge than a blanket grade for the whole conversation.
When you ask for feedback, ask for the statement concerned and an explanation of its effect on clarity. “You are too unsure of yourself” is a far-reaching interpretation. “Your sentence does not yet mention the transport question explicitly” is an observation you can work on in practice.
Say the agreement out loud yourself
To finish, put into words what is settled and what is still open. For example: “I’ll prepare the content. We’ll decide on transport after checking with the organising team.” In this exercise these are invented elements of an agreement, not statements about a conversation that actually took place.
Check whether the other role actually agreed or only mentioned a possibility. A fluent summary can unintentionally make a suggestion sound binding. That very difference is a useful thing to practise.
What the role-play cannot tell you
The AI role does not reliably reproduce how a real person would react. It is evidence neither of therapeutic success nor of objective conversation quality. Its answer can help you vary your own wording; it does not predict how the person concerned will react.
So take away a clearer opening and a useful follow-up question, not a supposedly finished outcome of the real conversation. There, listening and the actual answers remain decisive. The preparation helps if it lets you make your concern easier to understand while staying open to the other person.
Sources and documentation status: 25 October 2026.
