Here's the demo that gets everyone excited. You prop your phone against a rack, start your set, and a little skeleton lights up over your body on the screen. Knees drift in — the line flashes red and a voice says "knees out." You correct. Green. The machine coached you through the rep, live, the way a trainer standing next to you would.
It's a genuinely seductive pitch, and a big chunk of the AI-fitness category is built on it. It's also aimed at the wrong moment. Not because the technology is early — though it is — but because the rep is the wrong place to fix the rep.
Two problems, and only one of them is fixable
There are two things wrong with correcting form in real time, and it's worth separating them, because one is an engineering problem and the other isn't.
The first is that reading a moving body from a single camera, live, is genuinely hard. Depth is ambiguous. Limbs cross in front of each other. A barbell occludes the exact joints you care about. So live systems tend to fail in one of two annoying directions: they lag a beat behind the rep, or they false-flag — the line goes red on a rep that was actually fine, or stays green on one that wasn't. Anyone who's used one has had the experience of a coach in their pocket confidently correcting a lift that didn't need correcting. This is the problem people notice, and it's the one that might, eventually, get engineered away. Set it aside.
The second problem is the one that survives even a perfect system, and it's the one nobody's demo mentions: by the time anything — a camera, a mirror, a training partner, your own nervous system — can tell you a rep went wrong, the rep is already over.
You can't act on a cue you can't process in time
The mistakes worth catching don't announce themselves slowly. The knee cave, the lumbar flexion, the depth you lost all happen in a window of maybe 150 to 250 milliseconds at the hardest part of the lift. That's faster than a conscious reaction. By the time a red line appears on a screen and your brain reads it as "knees," the bottom of the squat is behind you. You didn't fix that rep. At best you fixed the next one — assuming you could glance at a screen mid-set without losing the brace that was holding your spine together.
And that's the quiet cost of the whole real-time idea: it's competing for the one resource a hard set actually needs, which is your attention. Under a heavy bar you have a limited budget of focus, and the lift wants all of it. Spending part of it monitoring a live readout, waiting to be told you're doing something wrong, doesn't make the rep better. It splits your attention at the exact moment the rep is most fragile. A coach who's any good knows this, which is why a good coach mostly shuts up during the set. They give you one cue before you start, and then they let you lift.
Form is a between-sets skill, not a during-the-rep one
So where does form actually improve? Not inside the rep. Between them.
Think about how a real coaching correction lands. It's almost never "no, right now, mid-lift." It's after the set: here's what I saw, here's the specific thing, here's what we're changing next time. You watch it back together, slow, from an angle you couldn't see from the inside. The fix gets encoded, and it shows up on the next set or the next session — as an adjusted setup, a different cue, a weight you weren't ready for yet. Correction is a review activity. It always was. The trainer's real work happens in the pause, not the pull.
Which means the useful thing a camera can do isn't yell at you in real time. It's the opposite: record the set cleanly, get out of your way while you lift, and then give you a calm, specific read of what actually happened — frame by frame, at half speed, pointing at the one thing that matters instead of ten things that don't. That's not a downgrade from real-time. It's the version that's actually pointed at the moment where form changes.
None of this makes a single camera all-seeing, and it's worth being honest that it isn't — one lens still can't measure how hard you're bracing or feel the load through a joint, and anyone claiming clinical certainty from a phone video is overselling. But a clean recording you can scrub back through is strictly more than you had, and it's paired with the thing that makes any of it stick: someone — or something — that remembers what it told you last time, so the correction compounds instead of resetting every session.
What this means for how you train
You can act on all of this today, with or without any app. Stop trying to fix reps while you're in them. Film your working sets, put the phone down, and give the lift your full attention. Then, after — not during — watch it back and look for one specific thing. That single habit will do more for your technique than any live overlay, because it's aimed at the moment where technique is actually editable.
It's also, deliberately, how we built Flexion. It doesn't try to argue with you between reps. You lift, you upload the set, and it reads the rep the way a coach reviewing footage would — frame by frame, calling out the specific fix and why it matters — then carries that forward, so next week it already knows what you were working on. The point was never a phone that talks during the set. It's the considered read afterward, from something that remembers.
The future of form coaching isn't a device narrating your squat in real time and hoping you can react faster than physics. It's the calm, honest look back — the part that was always where you actually got better.
