How to Make ChatGPT Your Personal Trainer: Prompts and Limits

ChatGPT works as a training reference that answers questions and fails as a coach that watches you, so the setup that works treats it as a programme writer whose output you supply the feedback for. Here are the prompts, the memory setup, and the three things to never take from it.

A disclosure that should shape how you read this: we build a nutrition database, not a fitness product. The training guidance below is general and widely published and we flag it as such. The food half is where we have actual expertise.

What can ChatGPT do as a personal trainer?

Split a trainer's job into four parts and it is good at exactly one and a half of them.

The first row is genuinely useful and free. The second is the one that matters most for injury risk. The third is possible with effort. The fourth is the reason most people hire a trainer in the first place, and it is the one an assistant cannot provide.

The setup that makes it workable

The failure mode is starting fresh every session. Fix that with a file you paste at the start, or a project instruction if your client supports one:

# Training profile
- 34, male, 82 kg, 180 cm. Two years lifting.
- 4 sessions a week, 45 minutes, full gym.
- Left shoulder dislikes overhead pressing. No pain otherwise.
- Current bests: squat 120x5, bench 85x5, deadlift 150x3.

## Rules
- Never prescribe a specific weight. Give rep ranges with an RPE target.
- When I give you session data, update the log, do not rewrite it.
- If I report pain, tell me to see a professional. Do not program around it.

## Session log

That third rule is deliberate. Assistants will happily suggest modifications for a painful joint, and that is the single clearest case for a physiotherapist instead.

Prompts that work

Writing the programme

Write a 4-day upper/lower split from my profile. No specific weights,
give rep ranges with RPE targets. State the progression rule and say
explicitly when it will stop working. Flag any exercise that commonly
causes trouble and give one swap for each.

After a session

Session done: [lifts, sets, reps, RPE]. Append to the log. Tell me what
to do next session for each lift, and flag anything that has not moved
in three sessions.

That second prompt is the closest thing to real coaching here, because you are supplying the feedback the assistant otherwise lacks. It only works if you actually record sessions.

Reviewing a block

Read the last six weeks of the session log. Which lifts progressed,
which stalled, and what single change would you make? Do not rewrite
the programme.

The last sentence prevents the most common failure, which is being handed a brand new plan every time you ask a question. Most results come from repeating something adequate long enough to progress on it.

Do custom GPTs help?

A little, and less than the marketing suggests.

A custom GPT is a saved set of instructions plus optional files. What it genuinely fixes is the re-pasting problem: your profile, rules and preferred output format persist, so you are not rebuilding context every session. That is real and it is most of the value.

What it does not fix is anything underneath. The model is the same, the training advice comes from the same place, and the food numbers are still recalled rather than looked up. A custom GPT labelled as a personal trainer is a prompt with a name.

The one thing worth adding to a custom GPT if you build one: a standing instruction to append to a session log and to refuse advice about pain. Behavioural rules survive better than content instructions.

A worked first month

Week one. Write the profile. Get a programme. Do not change anything, and record every session, including the ones that went badly.

Week two. Same programme. Start asking the after-session prompt so the log builds. Resist asking for a new plan.

Week three. First review. Ask which lifts have moved and which have not. Expect at least one stall, because that is normal and not evidence the programme is wrong.

Week four. Make one change, not five. The point of a log is being able to attribute a change to an outcome, and changing several things at once destroys that.

The discipline of one change per review is what separates this from generating a fresh plan whenever you get bored, which is the most common way people use an assistant for training and the least effective.

What does this cost compared with a real trainer?

Worth being straight, since it is the actual question behind the query.

An assistant costs a subscription or nothing and gives you programme structure, explanations and a searchable reference. A trainer costs meaningfully more per session and gives you technique correction, real-time adjustment and accountability, which are precisely the three things an assistant cannot do.

The sensible split for many people is a handful of in-person sessions to learn the main lifts properly, then an assistant for programming and questions afterwards. That puts the money where the risk is, which is technique under load, and uses the free thing for the part that is genuinely well documented.

Three things never to take from it

A starting weight. It is guessing about a body it cannot see.

Advice about pain. Any pain that persists is a question for a physiotherapist or doctor. An assistant will offer confident modifications and has no way to assess you.

Technique instruction as a substitute for learning the lift. Written cues are a supplement to seeing the movement done properly, not a replacement. Technique on loaded compound lifts is the largest injury risk in a gym.

A fourth worth adding: supplement recommendations. Ask and you will get a confident list. Supplement evidence is genuinely contested, the regulatory position differs by country, and interactions with medication are a medical question. This is one of the clearest cases where a pharmacist or doctor is the right source and a chat assistant is not.

How do you know it is working?

Decide the criteria before you begin, because judging after the fact is where wishful thinking lives.

For strength the signal is clear enough: are the main lifts moving over six to eight weeks, allowing for a bad week. For body composition it is slower and noisier, and measurements plus photographs across months beat a mirror on any given day.

What is not a signal: how hard a session felt, how sore you were afterwards, or how sophisticated the programme looks written down. Soreness is a particularly poor proxy, and chasing it is why people change programmes constantly and progress on none of them.

Put those criteria in your profile file so reviews are conducted against them rather than against how you happened to feel that week.

What about AI personal trainer apps?

A related question, since the alternative to prompting ChatGPT yourself is one of the many apps in this category.

They differ from a chat assistant in two useful ways. They hold your history, so the programme adapts without you pasting anything, and they usually have a defined progression engine rather than generating advice fresh each time. Both address exactly the weaknesses described above.

What they generally share with ChatGPT is the part that matters most for safety: they cannot see you either. An app that adjusts your squat load from logged reps is doing useful bookkeeping, not technique coaching.

The evaluation questions worth asking of any of them are the same ones: does it hold your history, does it have an actual progression rule you can read, can you export your data, and does the nutrition side use a real food database or generate numbers. That last one is the question most of them answer vaguely, and the answer determines whether your macro targets mean anything.

Does it work for beginners?

With one significant caveat, and it is the caveat this whole post turns on.

A beginner benefits most from the programme-writing half, because standard beginner programmes are simple, well documented and hard to get wrong. An assistant will produce a reasonable one.

A beginner is also the person for whom technique coaching matters most and who is least equipped to judge their own form. That is the half an assistant cannot provide at all, and reading cues while learning a barbell squat is not a substitute for someone watching you do it.

The practical resolution is the one in the cost section below: spend money on a few sessions to learn the lifts, then use the free thing for programming. Inverting that, which is what most people do, puts the spending where the risk is not.

The half you can actually verify

Every AI personal trainer setup comes with nutrition attached, and that is the half where confident wrong numbers hide.

A training plan cannot be checked against a database. Whether four sets of eight suits you is a judgement, not a lookup. But food composition is a data problem: how much protein is in 180 g of cooked chicken thigh has a right answer that lives in a table.

When ChatGPT tells you your post-training meal has 45 g of protein, it recalled that. Over a training block those errors compound in one direction, because the things people under-count are oils, dressings and anything eaten standing up. The mechanism is in can ChatGPT count calories.

The fix is a tool that queries a catalog, so the assistant searches and scales a real row instead of recalling one. That is this nutrition MCP server. It uses an API key header and is verified with Claude Code and Cursor, not inside ChatGPT, since browser sign-in is not enabled here. Tools in the docs, plan on MCP pricing. Eating to support training is covered in tracking a lean bulk, and a REST plan is the route if you are building software.

Before you start

If you are new to lifting, returning from injury, pregnant, or managing a medical condition, have a qualified human review the programme. This is not a disclaimer for form's sake: the cost of bad technique under load is measured in months.

What we have not measured

We have not run AI-written training programmes and measured strength or body composition outcomes, and we are not the right people to. No effectiveness claims, no comparison against human coaching, and no rankings of AI fitness apps appear here.

Related

• ChatGPT workout plan: what it gets right
• Can ChatGPT count calories?
• Tracking a lean bulk in Claude
• ChatGPT diet plan: what it does well