ChatGPT Workout Plan: What It Gets Right and What to Check

ChatGPT writes a structurally sound workout plan and cannot see you train, which makes it good at programme shape and unreliable at load, progression and anything involving your specific body. Here is what to take from it, what to ignore, and where the nutrition half can actually be verified.

One disclosure first, because it should change how you read this. We build a nutrition database, not a fitness product. That means we can be genuinely useful about the food side and we have no privileged expertise on programming. Everything below about training is general and widely published, and we say so rather than dressing it up.

Is ChatGPT good at writing workout plans?

At structure, yes. Training programmes are one of the most heavily written-about subjects on the internet, and the standard templates are well represented: upper and lower splits, push pull legs, full body three times a week, linear progression for beginners. Ask for a four-day split and you will get a sensible one with reasonable exercise selection and a defensible set and rep scheme.

At everything that depends on you, no. It cannot see your form, it does not know what you lifted last week unless you tell it, it has no idea that your left shoulder complains on overhead press, and it will not notice you are under-recovering. Those are the things a coach is actually for.

So the useful framing is that ChatGPT is a well-read training textbook that answers questions, not a coach.

What it gets right

Programme structure. Split, frequency, session order, sensible exercise pairings.

Explaining the reasoning. Ask why compound lifts come first, or why it put deadlifts away from squats, and you get an accurate explanation. Good for learning.

Adapting to constraints. Three days a week, only dumbbells, a bad knee, forty-five minute sessions. It handles several constraints at once, which is where template websites fail.

Variety without chaos. Ask for a swap for barbell rows and you get options that train the same pattern rather than something random.

What to check or ignore

Starting loads. Any specific weight it suggests is a guess about a person it cannot see. Use the structure, set the loads from your own recent performance, and if you have none, start deliberately light.

Progression rules. It will confidently give you "add 2.5 kg a week". That works for a while for a beginner and then it does not, and nothing in the chat notices when it stops working. Progression is the part of programming that most needs feedback, and feedback is the thing it does not have.

Volume claims. Specific weekly set counts get stated with more confidence than the evidence supports. Treat ranges as ranges.

Anything about pain. An assistant will offer advice on training around an injury. This is the clearest case for a physiotherapist or doctor instead. Pain is information about your body that no text model has access to.

Programme hopping. Ask for a new plan every fortnight and it will happily oblige. Most training results come from repeating something adequate long enough to progress on it.

A prompt that produces something usable

Write me a 4-day upper/lower split for someone who has lifted for two
years, 45 minutes a session, full commercial gym access.

Rules for your answer:
- Do not prescribe specific weights. Give a target rep range and an
  RPE or reps-in-reserve target instead.
- State the progression rule explicitly and say when it will stop working.
- Flag any exercise that commonly causes problems and give one swap.
- List what I need to track each session for this to be adjustable.

The first rule is the important one. Asking for RPE or reps in reserve rather than kilograms moves the load decision to where the information is, which is you, in the gym.

What does a good ChatGPT workout plan look like?

The output worth keeping has a recognisable shape. If what you get back does not look like this, ask again.

It names the split and the reason. Four-day upper and lower, because you said four days and forty-five minutes. Not a generic bodybuilding split pasted regardless of your constraints.

Compounds first, accessories after. Within a session, the lifts that demand most coordination come while you are fresh.

Rep ranges rather than weights. Three sets of six to eight at RPE 8, not three sets of six at 80 kg.

An explicit progression rule. What triggers adding weight, and what to do when you miss reps two sessions running.

Deload or reassessment built in. Programmes that only go up are programmes that stop working without telling you.

Swaps listed. One alternative per exercise, so a busy rack does not derail the session.

What should worry you: exact kilograms, a promise about how much you will gain, a plan that changes every week, or any advice about training through pain.

It should also tell you what it does not know. A good response asks what you can currently lift, or says plainly that it is assuming a starting point. One that produces a fully specified programme with loads, from nothing, has skipped the step where it should have asked.

Running, cardio and marathon plans

A common ask, and the answer differs from lifting in one important way.

Endurance programming is more formulaic than strength programming, so ChatGPT does relatively well at it. Ask for a twelve-week half marathon build and you get a sensible progression of long runs, easy mileage and a taper, because those structures are heavily published and fairly standardised.

The weakness is the same one: it cannot see your response to training. Endurance plans fail through accumulated fatigue and through raising mileage faster than tissue adapts, and neither is visible in a chat. The widely cited guidance about increasing weekly volume gradually exists precisely because that is where injuries come from.

Two practical notes. Ask for paces in terms of effort or heart rate zones rather than minutes per kilometre, for the same reason you ask for RPE rather than weights. And if it gives you a race prediction, treat it as entertainment.

Tracking it so the plan can adapt

A plan without a log is a plan that cannot improve, and this is the step people skip.

Session done: squat 100x5x3 at RPE 8, bench 75x5x3 at RPE 9,
rows 60x8x3. Append to the log. What should I do next session for
each lift, and flag anything that has not moved in three sessions.

This is where an assistant becomes something more than a template generator, because you are supplying the feedback it otherwise lacks. The flag on stalled lifts is the valuable part: humans are bad at noticing a lift has not moved in a month, and a text log is very good at it.

Keep the log in a file rather than the chat. Chats forget, files do not.

Where the numbers can actually be checked

Here is the part that connects to what we do, and the honest boundary around it.

A training plan cannot be verified against a database. There is no lookup that tells you whether four sets of eight is right for you. But the nutrition half of a fitness plan can be verified, because food composition is a data problem rather than a judgement problem.

If ChatGPT writes you a workout plan and a supporting diet, the diet is the half where confident wrong numbers hide. Ask it how much protein is in your post-training meal and you get a recalled figure, not a lookup. Over a training block that compounds. The mechanism is explained in can ChatGPT count calories.

The fix is to attach a tool that queries a food catalog, so the assistant searches, gets a row with macros per 100 g, and scales it to your gram weight. 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 are in the docs and the plan is on MCP pricing. Eating in a surplus while training is covered in tracking a lean bulk.

For software rather than personal use, a REST plan is the right route.

Is ChatGPT better than a free programme off the internet?

A fair question, since proven free programmes exist and have been run by millions of people.

A published beginner programme has one advantage no assistant can match: it has been tested at scale, its progression is known to work, and its failure points are documented by thousands of people who hit them. That is real evidence, and for a beginner it is hard to beat.

What an assistant adds is fitting. It will adapt a template to four days rather than three, to forty-five minutes rather than ninety, to dumbbells only, to a shoulder that dislikes overhead work. Proven programmes are rigid by design, and rigidity is a feature until it collides with your actual life.

A reasonable rule: if a well-known programme fits your schedule and equipment, run it. Use an assistant to understand it and to answer questions about it. Reach for a generated programme when your constraints make the standard options impractical.

How do you know if the plan is working?

Set the criteria before you start, because judging afterwards is where wishful thinking lives.

For strength, the signal is straightforward: are the lifts moving over six to eight weeks, accounting for a bad week here and there. For hypertrophy it is slower and noisier, and measurements plus photographs over months are more informative than a mirror on any given day.

What is not a signal: how hard a session felt, how sore you are afterwards, or whether the programme looks impressive written down. Soreness in particular is a poor proxy for anything useful, and chasing it leads people to change programmes constantly.

Write the criteria into your profile file so the assistant reviews against them rather than against how you feel that week.

Before you follow any of it

If you are new to lifting, have a medical condition, are pregnant, or are returning from injury, get a qualified human to look at the plan. Technique on loaded compound lifts is the single largest injury risk in a gym and it is the thing least amenable to written instruction from a system that cannot see you.

What we have not measured

We have not run training programmes generated by ChatGPT and measured outcomes, and we would not be the right people to. There are no strength or hypertrophy claims in this post, no comparison against coached programming, and no rankings of AI fitness tools.

What we can speak to with evidence is the food side, and even there we have published no accuracy benchmark yet.

Related

• Can ChatGPT count calories?
• Tracking a lean bulk in Claude
• ChatGPT diet plan: what it does well
• Nutrition MCP server: what the tools do