.png)
Ask an AI for a workout plan and you will get something that looks like a workout plan in about ten seconds. Sets, reps, rest periods, a tidy weekly schedule. The trouble is what you cannot see on the page. Because your prompt said nothing about your training history, your injury constraints, your equipment, or the days you can actually train, the model filled those gaps with an average plan for an average person. It reads well and it fails quietly. Two research teams who tested AI exercise prescriptions found exactly this pattern (Dergaa et al., 2024; Düking et al., 2024), and the second team found something more useful: the plans improved a lot when the AI was given more information about the person asking. The fix, in other words, is not better technology. It is a better prompt, and a better prompt has a learnable structure. This article walks through that structure and gives you three complete prompts you can adapt today, plus what to do when the AI gets things wrong and the three requests you should never make of it. For a broader look at this territory, we covered whether you can trust AI for health and fitness advice and where AI-driven personal wellness seems to be heading in earlier pieces; this one is about getting the most out of the tool anyway.
The default prompt, something like "make me a workout plan to get in shape," contains no information about you at all. A large language model (LLM), the type of AI behind chatbots like ChatGPT, is trained to predict likely text rather than to look up a single right answer, so it responds to an empty prompt with the statistical middle: a plan suited to a hypothetical healthy adult who trains regularly and has no constraints. You are probably not that person, and the plan will quietly assume you are.
Four missing inputs do most of the damage. Missing training history means the plan might assume you can load a barbell with confidence when you have never touched one, or assume you are a complete beginner when you are returning after years away. Missing injury constraints mean the plan might prescribe deep squats to a cranky knee, with no way of knowing the knee exists. Missing equipment means a program built around lat pulldowns and leg presses when your home gym is a pair of adjustable dumbbells. Missing days means a five-day split for someone with three free evenings, which usually ends as a two-day split by accident.
The output still looks good, and that is the trap. The text is fluent, organized, and specific-sounding, so it reads as expertise. When researchers asked coaching experts to rate ChatGPT-generated six-week running plans against twenty-two quality criteria, plans built on minimal input were rated below acceptable on nineteen of those criteria, and the plans improved significantly as more information about the runner was supplied (Düking et al., 2024). A separate expert evaluation of GPT-4 exercise prescriptions reached a similar conclusion from the other direction: the authors noted the prescriptions lacked a preliminary assessment, and wrote that if the AI had requested information such as the subject's current physical condition and exercise experience, it could have planned exercise in a more individualized and effective manner (Dergaa et al., 2024). Read that twice, because it is the whole article in one sentence. The researchers found the AI's biggest flaw was that it did not ask what you would have told it.
This matches a long-standing finding from health communication research. Noar, Benac, and Harris (2007) reviewed 57 studies of health behavior interventions and found that messages tailored to the individual outperformed generic messages. A prompt is your side of that tailoring job. The AI cannot individualize what it never heard about, so the missing context is not the model's failure. It is the part of the job you were quietly holding.
Every solid fitness prompt, whether it asks for a program, a food analysis, or a race plan, carries the same four parts. Context, constraints, output format, and a safety line.
Context is who you are and where you are starting from. Age, months or years of training experience, what you currently do, your goal, and honestly why you want it. "Beginner" is a start, but "sedentary for two years after a decade of recreational lifting" gives the model something to work with. If you have never assessed your starting point, a simple fitness self-assessment gives you numbers worth pasting into this section.
Constraints are the non-negotiables the plan has to work around. Injuries and the specific movements that provoke them, the equipment you actually own, the days and minutes you can realistically train, and any exercises you simply refuse to do. This section carries most of the safety value of the prompt, and it is the section people skip most often.
Output format tells the AI what shape the answer should take. A week-by-week table, sets and reps for every exercise, a progression rule, a rule for missed sessions, and one sentence of reasoning per choice. Format matters more than it sounds, because it forces the model to commit to a complete, checkable plan instead of a vague description of one. Asking for reasons gives you something to evaluate besides tone.
The safety line is a closing instruction that tells the model what to flag instead of bluff through. Something like: "Flag anything here that needs review by a doctor or physical therapist before I try it. If you are unsure about a guideline, say so rather than guessing." Models tend to answer confidently by default, so permission to express uncertainty has to be granted, and this is where you grant it.
Here is a complete prompt that follows the structure, filled in for a realistic example: a 43-year-old returning to exercise after two sedentary years, training at home.
Act as an experienced strength coach. I want a four-week beginner strength program.
CONTEXT: I am 43 and I have been sedentary for the past two years. I have
never followed a structured lifting program. My goal is to build a base of
strength and make training a habit I keep.
CONSTRAINTS: I have adjustable dumbbells from 5 to 40 lb, a bench, and one
resistance band. I can train Mondays, Wednesdays, and Saturdays for about
45 minutes in the morning. My left shoulder pinches when I press anything
overhead, so avoid that movement pattern.
OUTPUT FORMAT: Give me a week-by-week table showing every session with
exercises, sets, reps, and rest periods. Add one sentence per session
explaining why it is there. Tell me how to make each week slightly harder
than the last, how many warm-up sets to do, and exactly what to do if I
miss a session.
SAFETY LINE: Flag anything that a doctor or physical therapist should review
first. If you are not sure about a guideline, say so instead of guessing.
Compare what each part bought you. Three sessions a week is something you can check against expert guidance, since major sports-medicine guidelines recommend two to three training days per week for novices (American College of Sports Medicine, 2009), and a reasonable AI answer should land in that range. The dumbbell inventory means no barbell exercises appear on a plan written for a home with no barbell. The shoulder note pushes the model toward neutral-grip or landmine-style pressing instead of a barbell overhead press. The progression request is where progressive overload comes in, the practice of gradually making a workout harder over time so the body keeps adapting; we explain how progressive overload fixes flat-progress workouts in more depth elsewhere. Beginners also need a way to judge effort rather than only count reps, which our guide to measuring and monitoring exercise intensity covers without any gadgets. And the missed-session rule matters because life happens in week three, and a plan with no recovery path is a plan people abandon.
When the answer comes back, audit it against the prompt. Does every session actually fit in 45 minutes? Are all three days accounted for? Does anything sneak in an overhead press? A plan that ignores one of your stated constraints gets one follow-up message: "Session B includes X, which I said I cannot do. Replace it and reissue the table."
The second prompt flips the usual direction. Instead of asking an AI to write a meal plan you may not follow, you show it what you actually ate and ask for patterns. This is the diagnostic use, and it respects the fact that you already have eating habits and a life.
Act as a registered dietitian reviewing a food diary. You are analyzing,
not prescribing.
CONTEXT: Here is everything I ate for the last seven days, written as
honestly as I can manage: [paste seven days of rough notes, with portions
estimated as best you can]
CONSTRAINTS: Do not write me a meal plan and do not give me calorie
targets. I want to understand my own patterns first.
OUTPUT FORMAT: List the three or four patterns you notice, ranked by how
much you think each one affects my results, with one sentence of reasoning
for each. Then give me exactly one small change to try over the next two
weeks. Use plain language.
SAFETY LINE: If anything in this diary suggests I should talk to a doctor
or a registered dietitian, say so plainly. Do not diagnose anything.
Why analyze the past instead of plan the future? Because pattern-spotting is where AI is strong and measurement is where it is weak. Your portion estimates in a diary are genuinely rough, and they always are. Archer, Hand, and Blair (2013) examined self-reported energy intake in a large national nutrition survey and found that most records failed basic biological plausibility checks, meaning people of every description under-report or misreport what they eat. It should be noted that your diary has the same limits, so treat the AI's output as a list of observations about habits, not a verdict on your metabolism. What you want out of it is things like "you eat almost no protein before noon" or "your weekends run about double your weekday portions," ranked so you know which one to touch first. If protein at breakfast is one of the gaps, our guide to calculating optimal protein intake for muscle growth shows what a target looks like in grams rather than vibes. And the one-change rule is deliberate: behavior change research consistently favors small targeted adjustments over full dietary overhauls, and so does common sense.
The third prompt is the one with the most safety value, because it shows how to hand the AI a medical constraint without asking it to practice medicine.
Act as a strength coach who works alongside physical therapists. I want to
keep training my legs around a knee problem.
CONTEXT: I am 55 and I have lifted casually for three years. Three months
ago my doctor diagnosed patellofemoral pain (an ache under and around the
kneecap) in my right knee. My physical therapist cleared me for general
lower-body training and told me to avoid deep lunges and full-depth squats
for now.
CONSTRAINTS: I train at home on Tuesdays and Thursdays with dumbbells and
a squat rack. Deep lunges and full-depth back squats aggravate the knee.
OUTPUT FORMAT: Show me which movements to keep, which to modify, and which
to leave out for now, with a substitution for every exercise you are
dropping. Mark each exercise as "cleared" or "ask my physical therapist
first."
SAFETY LINE: This plan goes to my physical therapist before I follow it.
Flag every exercise that needs professional sign-off, and tell me what
warning signs during a session mean I should stop immediately.
The structure of this prompt is the point. Notice what it does not do: it does not ask "what is wrong with my knee" or "is squatting bad for knees," which are medical questions. It states the diagnosis and the clinician's instructions as facts, then asks the AI to do coach work inside those boundaries: keep, modify, replace. It also converts the AI from an authority into a drafter. The final plan goes to the physical therapist, and the prompt itself says so, which changes what kind of answer you get. Asking what warning signs mean stop is a question worth keeping even if you change everything else, because it turns the AI's knowledge of symptoms into something you can act on safely mid-session. If knee training interests you generally, we also looked at what the science actually says about the knees-over-toes squat debate, which touches this same territory from the research side. Nothing in this section is medical advice, and for any personal injury decision your physical therapist or doctor is the one to ask, not a chatbot.
AI models sometimes generate fluent, confident text that is simply wrong, and the research literature calls this hallucination: output that appears coherent but is factually incorrect or entirely fabricated (Ji et al., 2023). In fitness conversations this shows up as invented statistics ("studies show 73 percent of beginners quit by week six"), misattributed guidelines, and occasionally citations to studies that do not exist. Contradictions show up too: a plan that promises two rest days between hard leg sessions and then schedules them back to back, or a "beginner" plan that quietly adds a fifth training day by week three.
Four moves handle most of it. First, ask for the source: "which guideline is that number from, and what does the guideline actually say?" A correct answer gets more specific under pressure, while a hallucinated one tends to get vaguer. Second, ask it to reconcile: "your progression rule says add reps weekly, but week three cuts them. Which one applies?" Models respond well to being shown their own inconsistency. Third, start a fresh chat and re-run your four-part prompt from scratch, because errors and drift compound over a long conversation and a clean run is often the cheapest fix. Fourth, verify any named study before you repeat it to anyone, which is a good habit generally; our walkthrough on navigating scientific literature in health and fitness covers how to check whether a paper exists and what it actually found. If a citation fails your check, discount the whole answer until the rest of it is re-verified, because a model that invented one source may have invented others in the same paragraph. Spotting the shape of a bad claim is a skill on its own, and our free Truth in Fitness Marketing mini-course drills exactly that.
Do not ask an AI for a diagnosis. "What did I tear in my shoulder?" is a clinical question that requires imaging, a physical examination, and a license, none of which the chatbot has. The model will happily produce an answer that sounds exactly as authoritative as a radiologist's report, and there is no way to tell from the tone which one you got.
Do not ask it to make medication or supplement dosing decisions. Whether you can combine a painkiller with your prescription, or how much of a supplement to take, depends on your bloodwork, your kidney and liver function, and the full list of what you already take. That is prescriber territory, full stop. For a grounded look at what popular products actually do and do not do, our evidence-based guide to popular fitness supplements is a better starting point than any chatbot.
Do not bring emergencies or red-flag symptoms anywhere near a chat. Chest pain, sudden weakness or numbness, fainting, severe shortness of breath, or thoughts of harming yourself belong with emergency services or a clinician, immediately. An AI can wait; those cannot.
These three follow the same four-part structure on purpose, because the structure is the transferable skill. These three are samples; the full set, with the guardrails already written in, lives inside our flagship health and wellness education course, which also includes a module on AI literacy so you can brief and audit these tools yourself instead of hoping a prompt works. If you would rather start with something free first, the free fitness resources page and the free mini-courses library are useful starting points too.
Save the four-part structure as a note you reuse. Open your phone's notes app and write four headings: context, constraints, output format, safety line. Any fitness question you ever ask an AI gets pasted under those headings, in that order. Easier alternative: if four headings feel like a lot tonight, write just two, injuries and equipment. Those two alone eliminate most of the quiet failures.
Inventory your equipment and calendar before you prompt. Spend ten minutes listing what you actually own, with the weight range on your dumbbells, and the days you can genuinely train, not the days you wish you could. Easier alternative: photograph your home gym setup and describe the photo to the AI; it works with that.
Grade the plan before you follow it. When the answer arrives, ask two questions in follow-up: "why this exercise instead of X?" and "which guideline is that from?" Specific answers mean the plan deserves a trial week. Vague answers mean you treat the whole thing as a draft. Either way, check every session against the constraints you stated, because the AI will sometimes drop one anyway.
Analyse your food in the past tense. Log seven days of what you actually ate, as honestly as you can, then use the second worked prompt above. Easier alternative: three days beats zero, and a weekend day is the one you most want included, because that is where the patterns usually live.
Route anything medical to a clinician. Diagnoses, medication and supplement decisions, and red-flag symptoms are doctor questions, and the safety line in every prompt you write should remind the AI to send you there. That line is not decoration; it is the part that keeps the other three parts safe to use.
American College of Sports Medicine. (2009). American College of Sports Medicine position stand: Progression models in resistance training for healthy adults. Medicine & Science in Sports & Exercise, 41(3), 687–708. https://doi.org/10.1249/MSS.0b013e3181915670
Archer, E., Hand, G. A., & Blair, S. N. (2013). Validity of U.S. nutritional surveillance: National Health and Nutrition Examination Survey caloric energy intake data, 1971–2010. PLoS ONE, 8(10), e76632. https://doi.org/10.1371/journal.pone.0076632
Bull, F. C., Al-Ansari, S. S., Biddle, S., Borodulin, K., Buman, M. P., Cardon, G., Carty, C., Chaput, J.-P., Chastin, S., Chou, R., Dempsey, P. C., DiPietro, L., Ekelund, U., Firth, J., Friedenreich, C. M., Garcia, L., Gichu, M., Jago, R., Katzmarzyk, P. T., … Willumsen, J. F. (2020). World Health Organization 2020 guidelines on physical activity and sedentary behaviour. British Journal of Sports Medicine, 54(24), 1451–1462. https://doi.org/10.1136/bjsports-2020-102955
Dergaa, I., Ben Saad, H., El Omri, A., Glenn, J. M., Clark, C. C. T., Washif, J. A., Guelmami, N., Hammouda, O., Al-Horani, R. A., Reynoso-Sánchez, L. F., Romdhani, M., Paineiras-Domingos, L. L., Vancini, R. L., Taheri, M., Mataruna-Dos-Santos, L. J., Trabelsi, K., Chtourou, H., Zghibi, M., Eken, Ö., … Chamari, K. (2024). Using artificial intelligence for exercise prescription in personalised health promotion: A critical evaluation of OpenAI's GPT-4 model. Frontiers in Public Health, 12, Article 1304613. https://doi.org/10.3389/fpubh.2024.1304613
Düking, P., Sperlich, B., Voigt, L., Van Hooren, B., Zanini, M., & Zinner, C. (2024). ChatGPT generated training plans for runners are not rated optimal by coaching experts, but increase in quality with additional input information. Journal of Sports Science & Medicine, 23(1), 56–72. https://doi.org/10.52082/jssm.2024.56
Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Kim, Y., Ishii, E., Madotto, P., Xu, W., & Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), Article 248. https://doi.org/10.1145/3571730
Noar, S. M., Benac, C. N., & Harris, M. S. (2007). Does tailoring matter? Meta-analytic review of tailored print health behavior change interventions. Psychological Bulletin, 133(4), 673–693. https://doi.org/10.1037/0033-2909.133.4.673
Chris Bigelow holds a master’s degree in kinesiology from A.T. Still University and has taught kinesiology and exercise science at Bryan University for more than 10 years. He founded Innova Vita Fitness to make evidence-based health, fitness and research education more accessible to adult learners. His work focuses on exercise science, health literacy, sustainable behavior change and helping beginners evaluate health information more confidently.
Click Here
