The Prompt
Constraints first, role second. The constraints are what work.
Fill in the two bracketed lines and paste the rest as it is. The value is not in the opening sentence, it is in the six rules underneath it, which stop the model doing the three things it does by default: answering everything at once, reassuring you, and moving on.
Paste this into ChatGPT
You are acting as a reflective listener. You are not a therapist, not a doctor, and not qualified to diagnose or treat anything. Say so plainly if I ask you to be one. How to respond: - Ask one question at a time, then stop. Never stack three questions in a message. - Reflect back what you heard, in my words, before you respond to it. - Do not offer advice unless I ask for it twice. - Do not reassure me that it will be fine. Stay with the thing instead. - If I contradict myself, say so plainly and ask about it. - Keep replies under 120 words unless I ask for more. - If what I describe sounds like a crisis, stop the exercise and tell me to contact a real person or a crisis line. The situation: [two or three sentences about what is going on]. What I want out of this: [what a good outcome of the conversation looks like]. Start by asking me one question about it.
One thing to know before you start. This is a structured thinking exercise, not treatment, and the last rule in the prompt exists because the model needs to be told to break character. Read when to stop before you use it on anything heavy.
Why Most Therapist Prompts Fail
Role prompts get you the vocabulary of therapy and none of the method.
The common version is one line: act as a therapist and help me with my problem. What comes back is fluent, warm and useless, for four reasons that are all about the model rather than the wording.
It resolves instead of opening
A language model is optimised to produce a complete, satisfying answer to the message in front of it. Given a problem, it will summarise, empathise, list four options and close. That is the shape of a good answer and the opposite of a good session, where the first move is almost always another question.
It agrees with you
Models are tuned to be agreeable, so it will find the merit in whatever you have already decided. Being gently challenged by someone with no stake in your comfort is a large part of what you are actually looking for, and you will not get it unless you explicitly ask.
It reassures too early
Left alone it will tell you the situation is understandable and that you are doing better than you think. Premature reassurance ends the exploration. The rule about not reassuring you is the single most useful line in the prompt above.
It performs the role
Cast as a therapist, it generates therapy shaped language: validating openers, reframes, gentle summarising. It reads as competent and commits you to nothing. Describing the behaviour you want gets you much further than naming the profession.
Four Things No Prompt Can Fix
These are structural. No wording gets around them.
No duty of care. A licensed practitioner is accountable to a regulator, carries insurance, and can lose their livelihood for handling you badly. A model is accountable to nobody, and neither is the prompt you wrote.
No confidentiality. There is no privilege and no clinical record protection. Assume anything you type could be retained or reviewed, and leave out details that would harm you elsewhere.
No continuity. It does not know you between conversations unless you paste your own history back in. A practitioner remembers the thing you said eight months ago and notices you are saying it again, which is often where the work happens.
No stake in the outcome. It cannot notice you have quietly stopped showing up, cannot follow up, and will never be the reason you do the hard thing this week because someone is going to ask.
The honest summary
It is a notebook that asks questions back. That is genuinely useful, and it is not treatment.
When to Stop and Talk to a Person
The one part of this page that is not about prompting.
If you are thinking about harming yourself, if you cannot keep yourself safe, or if something has happened that you cannot carry on your own, close the chat and contact a person. A model cannot call anyone, cannot assess risk, and cannot stay with you.
United States
Call or text 988, the Suicide and Crisis Lifeline.
UK and Ireland
Call Samaritans free on 116 123, any time.
Anywhere else
Find a local line at findahelpline.com.
Immediate danger
Your local emergency number.
For anything persistent rather than acute, low mood that will not lift, a loss you are stuck in, patterns that keep repeating, the right next step is a licensed therapist rather than a better prompt.
Therapy, Coaching, and Which One You Want
A lot of people searching for AI therapy want the other one.
The two get used interchangeably and they are different jobs with different risk profiles. The distinction is worth two minutes because it changes what you should be looking for, and because it is the reason credible AI products in this space are built around coaching rather than therapy.
Therapy
Licensed and regulated. Treats mental health conditions. Often works backwards through your history to explain present distress. Assumes nothing about how well you are currently functioning, and is the right answer when something clinical is in play.
Coaching
Unregulated. Assumes you are functioning well and works forwards from where you are towards a goal you chose. No diagnosis, no treatment, and a responsible coach refers out the moment something clinical surfaces.
If what you want is to stop repeating something, or to understand why a thing still hurts, that is therapy and a prompt is not a substitute for it. If what you want is to get clear on a decision, hold yourself to something, or work out what you actually want next, that is coaching, and the exercise above is a thin version of it.
The fuller picture is in what coaching is and what a life coach does.
If You Are the Practitioner
The version of this question coaches ask.
A fair number of people who land on a page like this are not looking for support, they are coaches and consultants testing whether a prompt could stand in for them between sessions. It is a reasonable thing to test, and the answer is that a shared prompt gets you the generic version of the method while losing the two things clients pay for: your specific frameworks, and your judgement about when to use them.
That is the gap Personify fills. Instead of a prompt describing a generic listener, a clone is trained on your own material and answers in your approach, in 100+ languages, with you in the loop rather than a model improvising. The foundational questions get handled between sessions and your hours go to the work that needs a person.
Worth being explicit about the boundary: this is built for coaching, not therapy. If your practice is clinical, the accountability structure a licensed profession requires is not something a clone can carry.
Frequently Asked Questions
What is a good ChatGPT therapist prompt?+
A good one constrains the model rather than casting it in a role. Tell it to ask one question at a time and wait, to reflect your words back before responding, to withhold advice unless asked twice, and to keep replies short. Telling it to act as a therapist without those constraints produces therapy-flavoured language and very little listening, because the model will try to resolve your problem in a single reply.
Can ChatGPT be used as a therapist?+
No, and it will tell you so itself. It has no duty of care, no clinical training it is accountable for, no memory of you between conversations unless you supply it, and no confidentiality or privilege. It can be useful as a structured thinking aid, which is a genuinely different thing. Treat it as a notebook that asks questions back, not as treatment.
How do you prompt ChatGPT to be a therapist?+
Set the stance, set the rules of engagement, then give it the situation. State that it is a reflective listener rather than a clinician, list the response constraints such as one question at a time and no unsolicited advice, describe the situation in two or three sentences, say what a good outcome looks like, then ask it to open with a single question. The constraints do the work, not the job title.
Is it safe to tell ChatGPT personal things?+
Assume anything you type could be read by a human reviewer or retained. There is no therapist and patient privilege, no clinical record protection, and no legal confidentiality of the kind a licensed practitioner owes you. If a detail would harm you if it appeared somewhere else, leave it out or change it.
What is the difference between AI therapy and AI coaching?+
The distinction follows the human professions. Therapy is a licensed, regulated practice that treats mental health conditions and often works backwards through someone's history. Coaching is unregulated, assumes the person is functioning well, and works forwards towards a goal they chose. An AI that supports coaching sits inside a much smaller risk envelope, which is why credible AI coaching products keep a named human practitioner accountable for what the model says.
Will ChatGPT tell me what I want to hear?+
Usually, yes, and that is the largest practical problem with using it this way. It is trained to be agreeable and will validate a plan you have already decided on. Much of the value of talking to a skilled practitioner comes from being challenged by someone with no incentive to keep you comfortable. You can partly correct for this by explicitly asking it to argue the other side, but you have to remember to ask.
A prompt is generic. Your method is not.
Coaches use Personify to train an AI clone on their own frameworks, so the questions that repeat are answered in their approach and the hours go to the work that needs a human.
Free tier available. No credit card needed.