It makes up citations
Ask it for studies and it will give you some. With author, year and journal. Check three and one won't exist: the model looks nothing up, it predicts how a citation sounds.
Forge Coach
An AI coach that
cites its sources
It adjusts your diet and training every week by reading what you actually did. And when it decides something, it shows you what it leans on: the study, with its identifier, so you can check it yourself.
The coach is rolling out in stages. The app is free and works fully without it.
On track
Fairly confident
+0.8 cm on the thigh, −1.2 cm on the waist and steady weight: recomposition confirmed.
Calories
2,480 kcal
Protein
178 g
Quads
1417 sets
Backed by training.volume_landmarks · level B ·
Schoenfeld et al., PMID 27433992
Why this is different
Asking a language model for a plan always breaks in the same three places
It isn't that AI is bad at training. It's that a model on its own does three things that ruin a plan, and it does them with a confidence that fools you.
It does the maths differently each time
The same question, two weeks running, gives two different calorie targets. Not because you changed: because arithmetic isn't what it does. And a plan that contradicts itself gets abandoned.
It doesn't know what you did
You describe your week in a paragraph. It forgets the last one. Without your weigh-ins, your sets and your real log, any adjustment is a well-written guess.
Action and reaction
What it does, exactly, when something happens
Pick a situation. This isn't a marketing script: it's what the coach answers, using the same rules it carries inside.
You
I trained, but I logged no food at all.
- It leaves calories alone.
- It adjusts training, where it does have data.
- It tells you what was missing and what next week's review would be worth.
Changing calories without knowing what you ate is reacting to an incomplete signal. Weight trend only corrects the target when there's a log and minimum adherence.
You
The scale hasn't moved in three weeks, but I've lost 1.2 cm off my waist.
- It reads both signals together and concludes recomposition.
- It doesn't cut calories just because the scale is still.
- It may add sets for the muscle that's responding.
Tape measurements and weight trend lead the verdict because they're objective. A still scale with a shrinking waist isn't a plateau: it's exactly what you were after.
You
I did all 30 sets in the plan, but never logged how much I had left in any of them.
- It doesn't raise volume.
- It holds what you have and says so plainly.
- It asks for the effort on the last set of each exercise.
Without proximity to failure, thirty sets can be a huge stimulus or a walk in the park. Raising volume blind is the fastest way to pile up fatigue and gain nothing.
You
I ate out on Saturday and Sunday and went well over.
- You log it and it goes into the week's average.
- It doesn't tell you to make up for it on Monday.
- If the pattern repeats, it says so at the review.
Compensating turns one meal into a restrict-and-binge cycle. What decides the outcome is the average of weeks, not Saturday; and a plan that can't survive a Saturday isn't a plan.
You
I've gone more than ninety days without a period and I've been in a deficit for months.
- A clinical flag fires and the adjustment stops.
- It refers you to a healthcare professional, ahead of any number.
- It doesn't keep pushing the deficit.
Amenorrhoea under a sustained deficit is a sign of low energy availability. This decision is fixed in code and the model doesn't make it: it can't skip it no matter how well everything else is going.
You
I'm allergic to tree nuts.
- It suggests no recipe that contains them.
- Nor any that doesn't declare what it contains.
- It tells you how many were left out for not being labelled.
“Not stated” isn't “doesn't contain”. Suggesting an undeclared recipe would be claiming it's safe when nobody has checked, so when in doubt it stays out and you're told.
Where the decisions come from
A notebook of 53 rules that someone read one by one
The coach can only cite rules that exist in that notebook and are verified. If a rule isn't there, it doesn't exist for it. That's the only thing that genuinely stops it from inventing a study.
53
rules, every one with a DOI or PMID that resolves
6
level A: meta-analysis or solid consensus
40
level B: reasonable evidence, with caveats
7
level C: common practice or thin evidence
- More than two thirds had their wording changed on review. Not because they were made up, but because they claimed more than their source supports. Two ended up saying the opposite of how they were written.
- The level is set by the source you cite, not by the evidence that exists in the world. If all you can cite is a consensus, it's B even if the topic has meta-analyses.
- Strength and hypertrophy are kept apart. They respond differently to almost every training variable, and mixing them is the most common mistake in any gym advice you'll read.
What a rule looks like inside
cycle.water_retention · nivel B
“Fluid retention peaks at the onset of bleeding, not in the late luteal phase. It's about 0.5 kg of extracellular water, not one or two kilos.”
This rule was written wrong at first and corrected when it was checked. The app warned in the wrong phase and talked about kilos that weren't there.
Where the AI stops
The numbers don't come from the model
Calories, set distribution, weight trend and fatigue are computed by deterministic engines inside the app. The AI decides strategy and how to tell you about it; it never touches the arithmetic.
What the model decides
Whether this week calls for holding or moving something. What to explain and in what order. What to ask when the data doesn't arrive. When to send you to a professional.
What the app computes
Daily calories from your maintenance and your real weight trend. Sets per muscle. Accumulated fatigue. Streaks. All reproducible: same data, same result.
Why this matters so much
A coach that computes differently each week contradicts itself, and the moment it does you stop believing it. Taking the arithmetic out of the model isn't a technical limitation: it's what lets the plan still make sense three months from now.
The heart of it
The weekly review
Once a week it reads what you did — weight, tape measurements, logged food, sets and effort — and returns a verdict, an adjustment and next week's plan.
- Tape and weight trend lead. They're objective. The photo corroborates; it never decides.
- “I can't tell yet” is a valid verdict. With two weigh-ins and no measurements you don't claim a trend, and saying so is more useful than filling the gap.
- It shows you what was missing. Not as a telling-off: as the list of what would make the next review worth more.
- Photos with a ghost overlay. Last week's, translucent over the viewfinder, so you stand the same way.
Day to day
What you eat, without endless lists
The coach assigns you a pantry of recipes from Forge's catalogue that fit you, split by meal. You pick each day, and the list only changes at the weekly review, so you can actually do the shopping.
It never composes recipes with invented numbers
The recipes already exist and their macros come from their real ingredients. The coach picks among them; it doesn't conjure one up with calories eyeballed.
Allergies are a filter, never a preference
And a recipe that doesn't say what's in it is never suggested to someone who declared allergies. “Not stated” isn't “doesn't contain”: when in doubt, out.
The limits
What it doesn't do, said to your face
A list of limits sells worse than a list of promises. It's also the only way you'll trust what it does do.
- — It doesn't estimate your body fat from a photo. That's false precision and it shows. The photo corroborates what the tape and the scale say.
- — It doesn't replace a doctor or a clinical dietitian. There are situations where it stops and refers you, and that decision isn't the model's: it's fixed in code.
- — It doesn't programme your training by menstrual cycle phase. The evidence for doing so is weak. It does interpret fluid retention and flag red flags.
- — The catalogue recipes haven't been signed off by any professional. They're ordinary food with macros from composition tables, and the app says so on each one instead of calling them “validated”.
- — It doesn't guess what you don't log. The less you log, the less it claims. That's deliberate.
Frequently asked questions
What everyone tends to ask
How is this different from asking ChatGPT for a plan?
Three concrete things. First: the coach can only cite rules from a closed, verified notebook, so it couldn't invent a study even if it wanted to. Second: it computes nothing. Calories and volume come from deterministic engines in the app, so two weeks running with the same data give the same result.
And third, the one you feel most: it reads your real history — weigh-ins, measurements, sets, logged food — instead of a paragraph you type. When that data isn't there, it says so instead of filling it in.
Where does it get its information?
From a notebook of 53 rules checked one by one against the text of their sources, all with a DOI or PMID that resolves. Each rule carries its level of evidence, and that level is set by the source cited, not by what's been published on the topic.
On review, more than two thirds had to be rewritten: they weren't made up, but they claimed more than their study supported. Two ended up saying the opposite of how they'd been written.
What happens to my progress photos?
They're analysed and not stored on any server. All that's kept from the review is text describing what was seen. The copies that stay on your phone are there to help you stand the same way next week, and you can delete them whenever you want.
Consent is asked separately, before anything is sent.
Does it work if I don't log everything?
It works, but it claims less. With no logged food it leaves calories alone; without two measurements of the same girth the verdict rests on weight alone; without logged effort it won't raise volume.
And in every one of those cases it tells you what was missing, so the next review is worth more. That's the difference between a coach that admits what it doesn't know and one that fills the gap with something that sounds right.
What if I have an injury, a condition or take medication?
Clinical screening is done through a required form at the start, not in conversation, precisely so it doesn't depend on the model remembering to ask. There are situations where the coach stops and refers you to a professional, and that decision is fixed in code: the model can't skip it.
There are also clinical floors that correct the plan before you see it, and when they correct something you're told. A plan amended in silence isn't trustworthy.
Does it replace a personal trainer?
It replaces the part a good trainer does with a spreadsheet: reviewing your data every week, adjusting calories and volume, and justifying why. It doesn't replace someone watching you squat, or a healthcare professional.
What does it cost and how do I get it?
Forge is free and works fully without the coach: training, nutrition, calculators and progress. The coach sits in a higher tier and is rolling out in stages, so it isn't available to everyone yet.
Download the app and we'll tell you when it's your turn.
How often does the plan change?
Once a week, at the review. Deliberately: daily adjustments react to noise — water, salt, sleep, Tuesday's scale — and not to what's actually happening. Between reviews, what you've been assigned doesn't move, which is what lets you do the shopping and follow a plan.