I Handed My Running to an AI Coach. The Race Hasn't Happened Yet.

I've been half-skeptical of AI fitness coaching for a while now. Every app claims it'll unlock your true potential, and every claim falls apart the moment you ask for the actual numbers behind it. So when I set up a system where Claude builds my running plan from my Garmin data, I made myself a rule: no verdict until there's a real result to point to.

Here's the thing. There isn't one yet. My goal is a sub-48-minute 10k in Berlin, and I haven't run the race. So this isn't a "here's what AI coaching did for me" post. It's a mid-experiment log, written on purpose before I know how the story ends.

What the setup actually does

The mechanics are simple. Claude looks at my recent training in Garmin and generates the next block of structured workouts, each one tagged to a training week ("Wk1", "Wk2") with an explicit heart-rate target baked in.

An easy run means HR 133 to 143. A threshold session might be 3x8 minutes at a harder effort. Long runs sit at 14 to 16 kilometers, HR 137 to 147. Recovery days are recovery days, not an excuse to sneak in intensity. I still decide whether to run the session, skip it, or push back on it. The AI plans; I execute.

| Session type | HR target | Typical length | | --- | --- | --- | | Easy | 133-143 bpm | 6-10 km | | Threshold / tempo | above 143 bpm | 3x8 min intervals | | Long run | 137-147 bpm | 14-16 km | | Recovery | below 133 bpm | 4-6 km |

That structure is the whole product, really. No dashboard, no gamified streaks. Just a plan that knows what happened last week.

The numbers so far

From April 1 to July 12, I logged 58 runs, 507.6 kilometers, and 51.8 hours of running. That's not a huge volume by serious-runner standards, but it's a lot more consistent than what I managed before I let something else own the weekly structure.

Mid-block, I ran the Tromsø Midnight Sun Marathon on June 20: 42.74 km in 4:16:08, average heart rate 151. That wasn't part of the plan so much as a detour through it, a full marathon dropped into the middle of a block built around a completely different goal.

The interim signal: pace at a fixed heart rate has barely moved

This is the part I keep staring at. My easy-run pace at roughly 140 bpm is basically flat across the block so far. Late in the training, easy runs at HR 140 land around 6:06 to 6:13 per kilometer. Early on, at a similar heart rate, I was running 5:55 to 6:31 per kilometer, a wider spread landing in roughly the same range.

That's noisy data, and it's honestly not clean enough to call a trend either way. Summer heat pushes heart rate up at any given pace. Two work trips to Shanghai and Beijing broke up the routine. A handful of sessions happened on a treadmill, which runs differently than open road. I don't have a reliable VO2max delta to point to either, so I'm not going to manufacture one.

If you came here expecting a graph that goes up and to the right, I don't have it yet. What I have is a training log that's actually full, which counts for something on its own.

What three and a half months of AI coaching has actually given me so far

Consistency. That's the honest midpoint answer, and I think it's more useful than a fake fitness leap would be.

Before this setup, my running was reactive. I'd decide each morning whether to run and what to do, and "what to do" usually meant "whatever felt right," which in practice meant either nothing or the same easy 5k on repeat. Fifty-eight sessions with a defined job each time is a different thing entirely. The plan removed a decision I was bad at making for myself.

That tracks with how I think about AI tools generally: they're better at holding structure than they are at making you fundamentally different overnight. I made a similar point when I compared AI assistants against AI agents: the tool that quietly keeps state and hands you the next step is more useful, day to day, than the one promising a breakthrough. A Garmin plan generated by Claude is closer to the former.

It's also why I'm writing this before the race instead of after. I wrote about staying honest about limits in why AI content needs to admit what it can't do, and this is the same discipline applied to my legs instead of my writing. The plan holding context between sessions, remembering what I ran last week without me re-explaining it, is the same problem I described in context engineering for AI workflows: most of the value is just not starting from zero every time.

What settles it, and what doesn't

Race day settles it. If the AI-built plan gets me under 48 minutes for the 10k in Berlin, that's a data point worth something. If it doesn't, that's a data point too, and I'll write that one just as plainly. Either way, the honest answer today is that I don't know yet.

What I'd tell someone thinking about this in the meantime: don't expect a performance jump you can point to on a graph mid-block. Expect to actually run more often, because the daily "what should I do today" question gets answered before you open the app. That part already works. The rest is still ahead of me, and I'll report back once I've actually run the thing.