Can AI be your running coach?

Short answer: it can do one of a coach's three jobs very well, a second one with your help, and the third not at all. Which one depends less on the model than on what you let it read.

A running coach does three things. They write the plan. They review what you actually did against it and adjust. And they notice you: the limp you didn't mention, the week you're sleeping badly, the race you're secretly afraid of.

Without your data, an AI writes plans anyone could get

Ask ChatGPT or Claude for a 10K plan and you get a reasonable one: an easy base, a long run growing by ten percent a week, one session of intervals, a taper. It is the plan from every running book, because that is what the model has read. What it can't do is fit it to you. It doesn't know what you ran last week, at what heart rate, or how the last hard session went. So it asks you, and you guess, or it guesses for you.

That's not coaching. It's a template with your name on it. And the starting paces are where templates go most wrong, as the VO2 max trap shows.

With your data, it reviews like a coach

Connect the assistant to your Apple Watch workouts and the second job opens up. It can read a session rep by rep, and it can read months of them at once, which is where it beats most humans for patience. Here is a question I asked after a summer of training through heat waves:

Three things in that answer are what I'd expect from a good coach. It picked the right measure, distance per heartbeat on easy runs only, rather than average pace across everything. It raised the confounder before I did: summer heat raises heart rate at any pace. And it checked the confounder with the data instead of waving at it: September was still warm, and the gain was already there.

The same goes for a single session. Asked about an interval workout, an assistant that can see each rep's pace, power and heart rate will tell you that the slow rep was the wind, not your legs, because the power didn't drop. There's an example in ChatGPT vs Runna.

What it gets right

• Arithmetic across months. Weekly volume, the share of easy running, how your resting heart rate moved since spring: questions that take a spreadsheet and an evening take it ten seconds.
• Every rep, not the average. A 9 km interval session averages to nothing useful. Rep by rep, it says whether you held the target.
• All your sports at once. A hard week on the bike shows up in your runs. An assistant that reads both sees it; a running app doesn't.
• No ego. It doesn't defend the plan it wrote last week. Show it that the long run hurt and it changes the long run.

What it can't do

• See you run. It reads what the watch measures: ground contact time, stride length, vertical oscillation. It doesn't see your knee collapse inward.
• Know what you don't tell it. Pain, stress, a bad night the watch didn't record. Say it in the conversation; with HealthSync Pro, it can save a note on the workout so the next conversation remembers.
• Replace a professional for an injury. If something hurts in a way that changes how you run, that's a physiotherapist's job, not a chatbot's.
• Be right when the data is missing. An assistant asked about a workout it can't see will sometimes answer anyway. Ask it to quote the numbers it used; if it can't, it's guessing.

So, can AI be your running coach?

As a reviewer of your training, yes, and a better one than most of us are for ourselves. As the author of the plan, only once it can read what you've done. As the person who notices you, no: that part is still yours, and you have to tell it.

The setup that works for me: a plan from a running app, every session synced from the watch, and a question to Claude or ChatGPT after each hard one. If you're wondering what to ask, I wrote down five prompts that work.