The Algorithm Wrote Your Program. What A Generated Training Block Gets Right, And The Two Things It Cannot Know.
Asking a machine for a twelve week block is now a default feature rather than a novelty, and the output is better than most lifters expect. It is also competent in a specific and limited way. This article says plainly what generated programming does well, names the two things no model has access to, and gives you a way to use it that keeps the judgment where it belongs.
Opening. Nine seconds for twelve weeks
Tom is twenty seven and four years in. On a Sunday night he types his numbers, his four available days and his equipment into an app and asks it for a twelve week block.
It arrives in about nine seconds. It is formatted properly. It has an upper and lower split with sensible frequency, weekly set counts that sit inside ranges he recognises from his reading, a stated progression scheme, and a reduced week at week seven. There is a note explaining why the reduced week is there.
It is, honestly, better than the program he wrote for himself last year.
He runs it. Weeks one to four go well. Somewhere in week five something starts to feel wrong in a way he cannot name. The incline press is irritating the shoulder that gave him trouble two winters ago. The Friday session keeps not happening. His main lift has stopped moving while his accessory work keeps going up.
Nothing in the program is incorrect. It is a good program. It is just not a program for him, and the reason why is worth a whole article, because this is now how a large share of lifters get their training.
Section one. What the thing actually is
A generated program is the most plausible continuation of a prompt, produced by a model trained on an enormous volume of text. For training, that text is the published literature, the textbooks, two decades of coaching blogs and a vast quantity of forum argument.
So what comes back is a competent average of what the internet says about programming.
That is not an insult. The average of good coaching advice is considerably better than what most people write for themselves, and the model never forgets anything. It is also exactly why the output is never surprising, never specific to you, and never uncertain. A model has no mechanism for saying it does not have enough information, because producing an answer is the whole of its job.
Hold those two facts together. Competent and contextless. Everything below follows from them.
Section two. What it gets right, which is more than the sceptics admit
Structure. The honest research position on weekly volume, assembled in meta analyses by Brad Schoenfeld and colleagues, describes a dose response relationship between weekly sets per muscle group and hypertrophy with diminishing returns at the top end. A generated block will put you inside that range almost every time. Most self written programs do not.
Coverage. It will not give you five pressing movements and no rowing. It will not skip hamstrings for a whole block. A surprising number of self written programs do both, because we write programs around the lifts we enjoy.
Progression. It reproduces the standard models described in the American College of Sports Medicine position stand on progression in resistance training, and it will name which one it used.
Effort language. It will prescribe in reps in reserve, using the scale developed by Michael Zourdos and colleagues and published in 2016, which is a better instruction than a bare percentage.
Administration. It sequences, formats, counts and produces the thing in nine seconds. That is real value and it is unglamorous, which is why nobody writes about it.
If you have never followed a structured program, a generated one is better than nothing by a wide margin. Say that plainly before saying anything else.
Section three. The first thing it cannot know. Your history
The model has no access to your training history. If you paste a log it has what you typed, which is not the same thing.
It does not know that your left shoulder stopped tolerating flat benching in 2024 and has been fine on a slight incline ever since. It does not know that you have twice taken a reduced week and twice come back stronger than the block predicted. It does not know that your last plateau was food and sleep rather than volume, which is the diagnostic we walked through in The Plateau Decoded.
So it writes for the median lifter holding your stated numbers. The median lifter does not exist.
There is a quieter problem underneath this. A good coach asks five or six questions before writing anything, and some of those questions are uncomfortable. How many sessions did you actually complete last block. What hurts. What are you avoiding. The generator asks nothing, because it is built to produce rather than to interrogate. You receive confidence without context, which is the most expensive thing in training advice.
Section four. The second thing it cannot know. What the rep looked like
It cannot see you.
Execution is the variable that decides whether a prescription becomes training, and we made the full argument for that in Movement Quality Over Quantity. A set of eight at a given load is one thing performed well and a different thing performed with a bouncing lower back and a shortened range. The page cannot tell the difference. Neither can the app, however many sensors are attached to it.
The same hole sits under the effort instruction. It writes three sets of eight at two reps in reserve. Most lifters who believe they are at two reps in reserve are at four or five, which is the calibration problem The Last Three Reps was written to address. The model has prescribed the right thing and has no way to know the right thing did not happen.
And it cannot know what you will actually do. A program with four sessions you complete beats a better program with five sessions you do not. Adherence is the first variable in real training and the last one any generator can model, because it has never met you on a Friday.
Section five. Where the evidence actually sits
This is the part most coverage skips, so we will be plain about it.
There is no published randomised controlled trial comparing a generated training block against a coach written block in trained lifters across a meaningful period. Not a small one, not an inconclusive one. If such a trial exists we have not found it, and anyone telling you generated programming has been shown to match coaching should be asked for the citation.
What does exist is a small and early set of papers assessing whether exercise advice produced by these models agrees with published guidelines. The recurring finding in that literature is that the advice is broadly reasonable and contains specific errors, often around populations with medical considerations. That is a sensible thing to know and a long way from a performance claim.
There is also a question nobody asks their app. In Australia, health information is classed as sensitive information under the Privacy Act and the Australian Privacy Principles, which sets a higher bar for how it is collected and used. When you enter an injury history, a body mass, a sleep pattern or a menstrual cycle into a commercial fitness product, that is sensitive health information leaving your possession. Read what the product says it does with it. We made the structural version of this argument in Wearable Reality Check, and the cortisol wave is the same shape again. A measurement or an output you cannot interrogate is not knowledge. It is a claim with a good interface.
Section six. A defensible way to use it
Use it as a draughtsman. You are the architect. A draughtsman produces a clean drawing from your constraints, quickly, without errors of arithmetic. A draughtsman does not decide what the building is for.
Give it constraints rather than a question. Days available, session length, equipment, the exercises you know you execute well, the exercises that have caused you trouble, your current working loads, and what the block is for.
Ask for three versions, then choose one. The first answer is an average. Where three answers disagree is where a decision actually exists, and making that decision yourself is the entire point.
Audit before you run it. Four questions. Does every movement pattern appear. Is weekly volume per muscle group inside a range you can defend. Is there a planned reduction rather than an unplanned one. Does it progress on something nameable, so you can tell in week six whether it is working.
Keep the log yourself, in your own hand, including the sessions that did not happen and the reason. That log is the only thing in this entire system that is genuinely about you, and it is the one input no model can generate.
Hold the block for eight weeks. The strongest temptation with an instant generator is to regenerate whenever a session goes badly, which converts a program into a mood. Bad sessions are data, not a brief for a rewrite.
And a hard boundary. Pain, a diagnosis, a medication, pregnancy or a return from injury goes to a physiotherapist or a doctor. Not to a model, not to an app, and not to us.
Section seven. The mastermind frame
The part of training that a machine now does for free was never the hard part.
Writing a defensible block was always the easiest thing a coach did. The hard parts were the questions nobody wants asked, the honesty about what was skipped, the willingness to abandon a plan that was not working, and the patience to run something unremarkable for long enough to find out. None of that is a text generation problem.
The models will keep improving and they will keep improving at the same thing. Formatting, coverage, recall, speed. They will not acquire your history, because your history is not in the corpus. They will not acquire your eyes.
So let the machine draw the page. Bring the judgment, the log and the honesty about the Friday session. The body still only responds to what you did, and nothing has been invented that can do that part on your behalf.
Frequently asked questions
Is a generated program better than no program?
Almost always, yes. If the alternative is wandering a gym floor choosing exercises by mood, a generated block gives you coverage, structure and a progression you can audit. The honest comparison is not against nothing. It is against a program written by someone who has watched you lift.
Can I just give it my training log and solve the history problem?
Partly. It will use what you paste, which is better than nothing and still only the written record. It does not hold the pattern recognition a coach builds over months, it cannot see what the reps looked like, and it has no memory of you unless the product is explicitly built to keep one, which raises the privacy question rather than answering it.
Should I pay for an app that generates programs?
Only if the value is the admin, the reminders and the log, which are legitimate things to pay for. Do not pay for the generation itself, because that is now close to free. And read the privacy terms before entering an injury history.
Will it recommend supplements too?
It will, confidently, and that is where it is weakest, because the training literature is far better documented than the supplement literature and the model cannot tell you where the evidence thins out. Anything it tells you about a product deserves the same reading we give a label. What is the dose, what is the evidence, and who benefits from you believing it.
How will I know if the block is working?
Decide that before week one. Pick two or three measures, write them down, and judge at week six rather than week two. If you cannot name what would count as working, no program of any origin will tell you, and that is a problem with the brief rather than the draughtsman.
Written for the Supplement Superstore Vault. We sell the supplements that support the work. We do not sell the work. Information here is educational and is not medical advice. Speak with a qualified health professional before changing any protocol, especially during pregnancy, breastfeeding, competitive training or while managing any clinical condition.
