TL;DR
Give AI a client's age and nothing else and it writes for a demographic, not a person. Fix it with a four-part context block: training age and current numbers, joint history including what still works, how many hard sessions they recover from, and any medication that changes how you measure effort. Then state constraints rather than asking for swaps. Chronological age is the least useful number in the prompt.
Your new client is 56. She's been lifting for eleven years, pulls more off the floor than most of the men in her gym, and had a knee replaced two years ago.
You give the AI what you always give it: age, goal, training days, equipment.
What comes back is a program for a frail person. Three sets of ten on machines. Chair squats. A line about consulting her physician before beginning any exercise program. Somewhere in there the model read "56" and wrote for a demographic instead of a client.
Now flip it. Your other new client is 52 and has never trained in his life. The AI gives him five sets of back squats in week one, because he said he wants to get strong and nothing in the prompt suggested otherwise.
Both programs are wrong, and they're wrong in opposite directions. That's the tell. The model has nothing to judge with. You handed it one number and it filled the gap with an average.
The correction is the same one that fixes most AI programming problems: give it the things that determine the answer. For older clients, those things are specific, and they aren't the ones the model asks for.
What AI Gets Wrong About the 50+ Client
Four failure patterns show up again and again.
It treats age as a volume dial. Ask for a program for a 60-year-old and you'll get the 40-year-old program with fewer sets and lighter loads. Older clients often need a different distribution instead: similar hard work, spread differently across the week, with more attention to what happens between sessions than within them.
It confuses training age with age. An untrained 30-year-old and a trained 60-year-old are not the same problem, and neither are a trained 60-year-old and an untrained one. The model can't tell these apart unless you say so. Chronological age is the least useful number in the prompt, and it's usually the only one coaches give.
It plans recovery inside the session, not between sessions. AI is good at rest intervals. It's poor at spacing. What most often breaks a program for a client in their fifties or sixties is three demanding sessions stacked Monday, Tuesday, Wednesday, with no thought given to the fact that connective tissue takes longer to come back than muscle does. The program looks reasonable on paper and falls apart in week three.
It substitutes around joints in the wrong direction. Tell the model your client has a bad shoulder and it will swap the overhead press for a landmine press. Sometimes that's right. Sometimes the client's shoulder is fine overhead and hates the bottom of a bench press, and you've now removed the thing that worked and kept the thing that hurt. The model guessed which way the joint fails, because you didn't tell it.
AI is answering the question you asked with the information you gave.
The Context Block That Fixes Most of It
Before you write a single instruction, write four short chunks. This is the same context discipline behind the SCRIPT framework, applied to the details this population turns on.
1. Training age and current capacity. Not "experienced" or "beginner." How many years of consistent training, what they're currently doing, and one or two real numbers if you have them.
2. Joint and surgical history, including what still works. This is the chunk coaches get half right. Listing the bad knee isn't enough. Say which positions and ranges are fine, because that's what protects the exercises worth keeping.
3. Recovery reality. Sleep, work, stress, and how many genuinely hard sessions land in a week before quality drops. You know this number for your clients. The model has no way to guess it.
4. Medical and medication context, de-identified. Only what changes the programming. Some medications blunt heart-rate response, which makes heart rate a poor gauge of effort. Some make position changes uncomfortable. Keep names and documents out of the chat window. The de-identification habit matters more here than anywhere else, because this is exactly the client whose file has real medical history in it.
Filled in for the client from the opening:
Client Context Block
Client: F, 56.
Training age: 11 years consistent barbell training. Currently 4 days/week. Deadlift 275, squat 185, bench 115.
Joint history: Right total knee replacement 2 years ago, fully cleared, no restrictions from the surgeon. Deep flexion under load is uncomfortable past about 100 degrees. Loaded hinging, split stances and step-ups are all pain-free. Shoulders and back have no history.
Recovery: Sleeps 7 hours, works full-time, two grandchildren on Saturdays. Handles 3 hard sessions a week well. A fourth hard session shows up as a bad session, not a bad week.
Medical: On a beta blocker. Heart rate is not a usable effort gauge. Use RPE.
Goal: Keep her deadlift, add a little to the squat, stay durable.
That took two minutes to write, and it's reusable every block. Nobody reading it would write a chair-squat program, and nobody would write five sets of back squats either.
The Prompt
With the context block written, the instruction itself gets short.
Block Design Prompt
You are helping an experienced strength coach write a training block. Use the client profile below. Do not modify the program for age. Modify it for the specific constraints listed. This client trains hard.
[paste context block]
Write a 4-week block, 3 sessions per week. Requirements:
• Distribute hard work so no two demanding lower-body sessions fall on consecutive days.
• Progress load, not volume, unless you flag a reason.
• Every exercise must be compatible with the joint history above. If you're unsure whether something is, say so instead of guessing.
• Use RPE for effort, not heart rate or percentages.
• Include a deload trigger: tell me what I should see that means week 4 should be lighter.
Present it as a table by week and day. Then, separately, list every exercise choice you made specifically because of this client's constraints, and why.
That last line does more work than the rest of the prompt. Asking the model to justify its constraint-driven choices shows you where to look. When the reasoning is wrong, you'll see it immediately, and it's usually wrong in one or two places.
Substitution That Respects the Constraint
Exercise swaps are where most coaches use AI for older clients, and where the generic output does the most damage.
The problem is the shape of the request. "My client can't back squat, what should she do instead?" gives the model a hole to fill and no idea what shape it is. So it returns the list (leg press, goblet squat, split squat, hack squat) and you're back to choosing from a menu you could have written yourself.
State the constraint instead of asking for the swap:
Substitution Prompt
My client can't tolerate deep knee flexion under load past roughly 100 degrees. Everything else in the lower body is available: hinging, split stances, step-ups, and loaded carries are all pain-free, and she can load them heavily.
I need a knee-dominant movement she can progress for 8 weeks and take genuinely heavy. Give me three options ranked by how well they let her keep loading, and tell me what each one gives up compared to a back squat.
Now the model is solving your problem rather than a category. The ranking and the trade-off question matter: you want to know what you're losing, because with a client who's been training eleven years, you're often choosing between two imperfect options rather than finding a clean replacement.
The same move works for shoulders, backs, hips, and everything else. Describe the boundary and what remains inside it. Never name the body part alone.
What to Check Before It Goes Out
The errors AI makes in programs apply here too, but three checks are specific to this client.
Count the hard days and where they land. Count the demanding sessions (the total is not the number that matters) and look at the calendar spacing between them. This is the most common failure and the easiest to spot.
Read every exercise against the joint history you wrote, not the injury label. You listed which positions work. Check the program against that line, not against the word "knee."
Look for the accessory drift. Programs for older clients tend to fill up with balance work, band exercises and mobility circuits that nobody asked for. Some of it is worthwhile. Most of it displaces the loading that's doing the actual work. If your client is strong and wants to stay strong, three sets of clamshells are a poor trade for a set of heavy carries.
Five minutes, and you've caught the things that would have cost you a conversation in week two.
Where You Still Have to Decide
Some of this doesn't delegate.
Clearance is not a programming question. New symptoms, recent surgery, chest pain, anything cardiac, uncontrolled blood pressure: that's a physician conversation, and no amount of context in a prompt substitutes for it. The model will happily program around a description of a medical condition it has no business interpreting. Same boundary as working outside your scope on nutrition: the tool doesn't change what you're qualified to decide.
Pain quality is a judgment call. The difference between "this is uncomfortable and fine" and "stop" is a coaching read, made in person or on video, in the moment. You can tell AI the outcome of that read. You can't ask it to make it.
Whether this client is a masters athlete or a deconditioned adult is your call, not the model's. It's the decision the rest of the program hangs on, and it comes from knowing the person. Get it right in the context block and everything downstream follows. Get it wrong and no prompt engineering saves the program.
Here's the thing about coaching older clients well: most of what makes it work is knowing precisely what this person can and can't do, and being unwilling to round that off to an age bracket. AI will round off every time you let it. The context block is how you stop it.
Get the Prompts That Do This
The SCRIPT Toolkit includes the context templates and programming prompts behind this workflow: the ones for constraint-driven substitution, block design, and the review pass that catches what AI misses. $39 founders price for the first 100 buyers, then $59.
Get the Toolkit →Frequently Asked Questions
Does AI actually know how to program for older clients?
It knows the general principles, and it applies them too broadly. Ask about training after 50 and you'll get sound, generic guidance about recovery, connective tissue and progressive loading. The gap is that the model defaults to the average older adult unless you describe the one in front of you. A trained 58-year-old and an untrained 58-year-old need different programs, and only you can tell the model which one you have.
What should I include in a prompt for a client over 50?
Four things: training age and current numbers, joint and surgical history including which positions still work, how many hard sessions they recover from in a week, and any medication that changes how you measure effort. Chronological age matters least. Write these once per client, reuse them every block, and update them when something changes.
How do I stop AI from making the program too easy?
Say so directly, and give it evidence. "This client trains hard" plus real numbers from their current program does most of the work. If the output still arrives soft, the fastest correction is to point at what they already do: "She currently deadlifts 275 for triples. Rewrite this at a difficulty that makes sense next to that."
Is it safe to put a client's medical history into ChatGPT or Claude?
Not in raw form. Summarize instead of pasting, use initials or "the client" rather than names, and keep physician letters and imaging reports out of chat windows entirely. Include only what changes the programming. "On a beta blocker, use RPE not heart rate" tells the model everything it needs without handing over a medical record.
What about clients over 70?
The workflow doesn't change, but the context block carries more weight and the clearance question comes up more often. Balance, fall risk, and daily function usually move up the priority list ahead of maximal strength, and that's a decision you make and state in the prompt rather than one you let the model infer.