My wife could have taught me this in an afternoon. She coaches kettlebells. She can watch someone swing for three seconds and tell them what their hips are doing and whether their spine is safe. I decided to learn from an app instead, one I built myself, partly to see whether the software could do what she does. It could not, and she was the one who saw it first. The part worth writing down is exactly what she was right about.
The test was simple. Learn a dangerous physical skill from a screen, with no trained eye anywhere in the process, and see how far the screen could take me. I wanted to know whether the tool had gotten good enough to make the expert optional. I had the expert. I chose the screen on purpose.
The app is good. Seven movements, seven reps, seven rounds. A four-week on-ramp that will not let me skip ahead. There is a classroom with short lessons and quizzes, drills that isolate one piece of a movement at a time, and a film-and-review step where I record a set, watch it back, and grade my own form against a checklist. It gathers detail beautifully. It films the rep, it tells me exactly where to look, it hands me the short list of things that separate a clean swing from one that hurts me.
The swing is the one that hurts you. The rest of the seven are comparatively forgiving. The swing is a ballistic load and a hinge at speed, and the failure mode is a lower back that rounds under it. That rounding is not a figure of speech. It is a spine doing something a spine should not do while a weight is accelerating away from it. The margin is small; the difference between a hinge and a squat, between load through the hips and load through the spine, is a few degrees you cannot feel when you are new. People herniate discs this way. You find out you got it wrong the next day, in your back, when it is too late to take the rep back. The mechanism is real and it is close.
I made one decision early that I still think was right. The app does not judge my form. No computer vision, no pose estimation, no red line drawn across my lumbar spine. I wanted to build the habit of looking, not hand the looking to a model that might be wrong in ways I could never check. What it gives me instead is a lens. It tells me where to look. On the swing it says watch the bottom, watch whether the back stays flat, watch whether the power comes from the hips. Then it hands the call back to me.
That is where it fell apart, and the flaw is embarrassing once you see it. The review screen ends with three buttons. Pass. Fail. Unsure. I record the swing, I watch it, I tap one to log the rep. I did this for a couple of weeks, feeling responsible and thorough, before I sat with what I was actually doing. A beginner cannot confirm a swing. I do not have the eye. I cannot see a fault I have no concept for. The rounding a trained person catches in half a second is, to me, just a rep that felt fine. The button asked me to certify something I was in no position to certify. I had built a machine that gathers detail and called the act of rubber-stamping its output validation.
Underneath it was a confusion about what validation is. Validation is two separate jobs. There is gathering the detail, and there is judging it. The app is superb at the first and silent on the second, which was supposed to be fine, because the judging was meant to be the human’s part. The trap is that judging is the exact thing a beginner does not have. The software had reproduced the easy half, the collecting, and left the hard half sitting with the one person in the experiment who could not do it.
Karla made the point after I had already built the thing. I showed her what I had, pleased with it, and she did not think much of it. Her argument, as I remember it, was that this is not a skill you learn safely from a screen; the swing needs a trained eye on it, and that eye is the whole point. She did not need to watch me train to say so. She heard the design and knew what was missing from it, which is the same move she makes on someone’s hips in three seconds, pointed at a different object. I disagreed and kept using the app. It made her case for her the first time I sat in front of a review screen and understood that I had no idea whether the rep I was about to grade was safe.
What she was right about was not the swing at all. It was the structure of the whole thing. The eye is the part that matters, and the eye is the part I had tried to skip. You cannot manufacture it by watching more footage of yourself; you get it from a trained person looking at you, again and again, until the pattern loads into your own perception. The software can hold the checklist. It cannot hold the eye. Gathering the detail is the half that scales, because a machine never tires and never cuts a corner. Judgment is the half that does not scale, because it lives in a small number of trained people and it costs their attention to spend. The tool was never going to replace her. The most it can do is make the record so complete that one look from her settles it. It lowers the cost of reaching the eye. It does not build one.
None of this is abstract while I keep grading my own reps. A false pass does not cost me one bad swing. It grooves the pattern. My nervous system takes the rep I just certified and files it as the reference. I do it again, and again, through four more weeks of an on-ramp built to drill that exact repetition. The bad rep becomes the baseline. By the time I have the eye to catch the fault, the fault is no longer something I am doing; it is how I swing. The error stops being an event and becomes the record, and everything after it inherits the record.
Most of this does not need a trained eye, and the argument falls apart if I pretend otherwise. Nobody should be sending the rest timer out for expert review. The app might call for ninety seconds and give me eighty-five, and the error is cheap, obvious, and self-correcting, and none of it gets written into how I move. Judgment is worth reaching for in three places. Where the error is expensive. Where it cannot be undone. Where it gets absorbed into memory and shapes what comes next. Everywhere else a review is ceremony, and ceremony crowds out the checks that matter. Naming what I am glad to let slide is what earns the right to be stubborn about the swing.
The app is still on my phone. Its job just changed. It is a camera and a checklist now, a way to package a clean rep for the person who can actually read it, and I have stopped pretending the button means anything. I still film my swings. I still cannot see the third one for what it is. Karla can, in a few seconds, the way she always could. She was right about the one thing I built the whole app to disprove, and the app turned out to be the fastest way I have to hand her the question.