CommunityPricing

How it works

OverviewFind a practice partnerSchedule and callTranscripts and feedbackLogbookExam practiceSession recordingRobin’s feedbackAI Practice

Who it’s for

For coachesFor coaching schoolsFor coaching communities

Resources

GuidesBlogHelp centreFAQ
← All guides

Last checked 11 September 2026

Guide9 min read

How to practise coaching with ChatGPT as your client

You coach. ChatGPT plays a fictional client. State those roles before you start, give it a bounded scenario, and end the roleplay before asking for a debrief.

You coach. ChatGPT plays a fictional client. State those roles before you start, give it a bounded scenario, and end the roleplay before asking for a debrief.

You can repeat a fictional conversation without coordinating a partner's diary. Repetition does not establish improved coaching or supply real-client experience for a credential. Privacy risks and the possibility of rehearsing unhelpful habits remain.

A 2025 counselling study used a similar procedure. Trainees were asked to open ChatGPT, paste a standard roleplay prompt instructing it “to act as a client named Omar”, conduct a 30-to-40-minute text counselling session, then submit the transcript for faculty feedback. Nine trainees submitted transcripts and four joined a focus group afterwards. They were training as counsellors, not professional coaches, and the study did not measure skill change (Learning Through Simulation, 2025). Their experiences inform some of the suggestions below; they do not validate this coaching prompt or today's ChatGPT behaviour.

Decide what the conversation is for

Pick one thing to work on before you pick a scenario. “Get better at coaching” gives you nothing specific to notice afterwards. “Ask one question at a time”, “check what the client wants from the session before I go looking for solutions”, or “stay with the topic they raised rather than the one I find interesting” give you something to look for in your own responses.

One focus per rehearsal is enough. You can run the same scenario again with a different focus. Keep this exercise to an ordinary, low-stakes work or life decision; leave clinical, crisis and specialist-advice scenarios out of it.

Sort the settings before you sort the prompt

Use invented details throughout. Don't lightly disguise someone you coach, and don't use a friend's situation because it's handy. Leave client notes, real transcripts and other people's files out of the chat.

OpenAI's Data Controls FAQ says that turning off Improve the model for everyone applies across your signed-in account. New ordinary conversations can remain in history but won't be used to train ChatGPT. Check the setting before beginning.

For a rehearsal, consider a non-personalised Temporary Chat. OpenAI says it does not use memory, custom instructions or plugins, and does not create new memories. A personalised Temporary Chat can use existing memories and custom instructions. Paste the role and scenario directly into the new conversation rather than relying on saved instructions (Temporary Chat FAQ).

Temporary Chats do not appear in history or train the models while they remain temporary. OpenAI may keep a copy for up to 30 days for safety. Saving one converts it to a regular chat governed by your account's personalisation and model-improvement settings. OpenAI also says limited safety-relevant context from prior conversations may still be used despite turning off personalisation. These controls do not promise a completely isolated or risk-free conversation.

If you choose voice, check the ChatGPT Voice FAQ and the options available in your account. Transcripts may be used for training depending on your plan and settings when Improve the model for everyone is on. Audio and video clips are not used for training unless you opt to share them.

Retention differs by voice mode. OpenAI says Live and Advanced audio clips, and Advanced video clips, are retained for 30 days. Deleting the chat deletes associated clips within 30 days, with security, safety and legal exceptions, and an exception for previously shared training clips already disassociated from your account. Standard voice audio is deleted after transcription unless you chose to share it for training. Voice transcripts may not exactly match what was said.

These are official documentation descriptions checked on 11 September 2026, not observations from a signed-in account. Check your current settings; features and labels may differ.

Write a brief, not a sentence

Trainees described benefits from writing more than a line of setup. One said: “The more that I prime the situation, the more that I give complexities for it to build into the scenario, the more specific my feedback is.” Another comment described the cost: “If you go so far as to make it that helpful, then you're writing a whole biography before you even start.” These are participant reports, not a measured comparison of prompts (study, section 3.3).

Try a brief that names who the fictional client is, what they came with, what they are not ready to say yet, and how they should respond when a question misses. This also follows OpenAI's general advice to give clear, specific instructions and refine them after reviewing the response (prompt guidance).

The study's authors recommend asking for descriptions of non-verbal behaviour and telling the model to avoid being overly agreeable. Some transcripts included “[fidgets with hands]” and “[takes a deep breath]”; others omitted those cues. One participant described the AI as “the world's most pleasant client” (study, sections 3.4 and 5). Written stage directions are simulated cues, not observation of a person.

An illustrative starter prompt

This prompt has not been run against the current ChatGPT product for this guide. Treat it as a starting point to adapt, not a tested recipe. Nadia and her circumstances are fictional.

You are playing a fictional coaching client so that I can practise coaching. I
am the coach. Stay in the client role for the conversation. Do not coach me or
suggest what I should say. Step out of role only when I type PAUSE ROLEPLAY or
STOP ROLEPLAY. PAUSE is for adjusting the simulation; resume when I type RESUME.
STOP ends the roleplay before a separate debrief.

The client: Nadia, 34, a team leader in a housing association. She was asked
last week whether she wants to take on a second team. She has not decided. She
says she is "just being sensible about it" and is quicker to talk about her
current team's workload than about what she wants.

How to behave:
- Open with one or two sentences only. Let me ask before you explain more.
- Hold something back. Nadia is not ready to say that she resented how the
  offer was made.
- Do not agree automatically. If a question misses, answer the question I
  asked, disagree where it fits the brief, or say you're not sure what I mean.
- Do not force a quick resolution or thank me automatically for insights.
- Keep Nadia's history consistent with this brief and your earlier answers.
- In text, include occasional non-verbal behaviour in square brackets, for
  example [pauses] or [looks away].
- If I type "silence", treat it as me waiting rather than as an empty message.
- Keep the situation within this ordinary workplace decision.

Begin when I say hello.

The study recommends typing “silence” so the model interprets it as a pause. That lets you rehearse choosing to wait; it does not reproduce sitting through a real-time silence with a person. The prompt cannot guarantee how ChatGPT will respond.

Coaching it

Ask a question, read the answer, and let the answer inform the next question. Notice if you start working through a prepared sequence regardless of what the fictional client says.

Text gives you time to think before sending a response. The study's authors recommend starting with text for counsellors at the pre-practicum stage; that recommendation has not been tested here with professional coaches. Voice offers spoken turn-taking, but its transcript may not be an exact record. Choose the format for what you want to rehearse, rather than treating voice as a required next level.

When it drifts

Watch for these possible problems. The recovery lines are illustrative instructions, not observed results.

It agrees with everything. Pause the simulation: “PAUSE ROLEPLAY. Nadia is accepting every suggestion. Keep her uncertainty and let her disagree where it fits the brief. RESUME.” Try adjusting the brief, but persistent agreeableness may be a model limitation. Making the fictional client argumentative would create a different distortion.

It resolves too fast. One trainee said: “In just a few prompts, I solved all their problems… like they're ready to go and do all the homework” (study, section 3.4). A scripted breakthrough is not evidence that your question helped a person. You could restart with the instruction that Nadia leaves the session still undecided.

It starts coaching you. Try: “PAUSE ROLEPLAY. I am the coach. You are Nadia, the client. Answer my last question from her perspective. RESUME.” If it keeps reversing roles, end the chat and start again with the brief. The 2025 study excluded role-confusion coding because its authors judged that earlier problems had largely been resolved; it did not establish that current ChatGPT will stay in role.

It changes its story. Pause and name the contradiction: “PAUSE ROLEPLAY. The brief says Nadia hasn't decided. Your last answer says she already accepted. Return to the undecided situation. RESUME.” If the history remains inconsistent, restart rather than treating the inconsistency as something to interpret about the client.

You can stop an unhelpful rehearsal. An odd answer is not a fact about how people work.

Stop the role, then debrief

Type STOP ROLEPLAY before asking about your coaching. Do your own reflection first: what were you trying to do, where did you notice yourself drifting, and what would you try differently?

If you want the model's observations, ask for them tied to specific turns rather than a general verdict. For example: “Look at where I asked more than one question in a turn. Quote my exact words, and separate what you can see in the text from your interpretation.” This debrief instruction is also untested. Check every quotation against the text exchange yourself; voice transcripts have the additional accuracy limits described above. Discard feedback that misquotes the session.

Describing apparent insight from the AI client, one trainee said: “I may have viewed it… as validating me. I felt like it was more a part of its agreeableness, as opposed to a client's self-awareness” (study, section 3.5). Model-generated praise is not an assessor's judgement or feedback from someone you coached.

Take one observation to human practice

If you want to try one observation with a person, choose a consenting adult who knows you are practising. Keep the topic low-stakes, agree confidentiality limits and either person's right to stop, and obtain separate consent before recording. Clinical, high-risk and specialist-advice work are outside this exercise. A human partner can tell you what your questions felt like from their perspective.

Keep fictional rehearsal separate from client hours for a credential. ICF defines client coaching in terms of someone who has hired you specifically as their coach and real-time interaction; its experience guidance does not establish simulated-client eligibility (ICF experience requirements). If you record this exercise, label it fictional rehearsal, not client experience. Check your own body or programme's rules rather than assuming it fulfils a development requirement either.

For help arranging a human conversation, read how to find coaching practice partners and clients.

Sources

Relevant passages checked on 11 September 2026 through Jina after direct access was blocked. This is source-content review, not a current ChatGPT UI or prompt test, and does not establish uncached origin freshness.