The best coach I ever hired can't hold a racket

I have played tennis for over fifteen years. Long enough to know exactly which parts of my game were stuck, and long enough to have stopped believing they would unstick themselves.

What changed was not a new racket or a new club. It was that I started treating my tennis the way I treat my work: film it, feed it to a model, get the analysis, drill the fix, ship the next match.

The loop looks like this. I play a match and record it. The video goes into Gemini, which comes back with a detailed read of the match: patterns I repeat, points I give away, where I stand when I should be somewhere else. Between matches I photograph my shots, the preparation, the contact point, the finish, and I put the frames in front of ChatGPT and Claude and argue with them about technique. Before the next match I do the same for strategy: who I am playing, what I know about them, what the plan is when the plan stops working. One of those matchups was specific: a player who lives on drop shots and a heavy forehand. I argued it out with Claude and ChatGPT beforehand and walked on court with the plan already built.

The first catches were embarrassingly physical. Gemini watched my match footage and said it plainly: bend the legs more, stretch toward the ball, stop arriving at my own standing height. Then it read my serve at up to 160 km/h and handed me the homework every club player skips: the trophy position. My serve now has a number and a to-do list. Both came from a model.

None of the models has ever held a racket. That turns out not to matter. What a coach mostly sells is attention and a feedback loop, and attention is exactly what a model has infinite amounts of. It watches the whole match. It does not get bored in the second set. It does not soften the verdict because we are friends.

The feedback loop used to be measured in months. A club coach sees you one hour a week and remembers half of it. Now the loop closes in a day. Play Saturday, analysis Saturday night, drills Sunday, different player by the next Saturday.

There are things the loop cannot do. It cannot make me fitter. It cannot play the nervy points for me. It cannot force my feet to move when my brain has decided the ball is unreachable. The models are the coach, not the player. The gap between knowing and doing is still mine to close, and it closes the boring way, with repetitions.

What strikes me most is that this is the same loop I run at work. Write the spec, route it to the model, verify the output, ship, keep what worked in the library. My serve gets the same operating system as my dashboards. The whole site you are reading was built with it.

Fifteen years of tennis taught me how to hit the ball. The models are teaching me what I actually do when I hit it. Those turned out to be two different educations, and I needed both.