About a mile from the finish on Sunday, I stopped to take in a gel and some water. And my stomach turned over.
I was seven and a half miles around Mackinac Island, in a wetsuit, with nothing to do about it. So I breathed. Slow and deliberate, for longer than I wanted to. It took a while to settle.
Nothing in my training plan covered that. My coach was on a laptop back home.
I have now done the 8.2-mile swim around Mackinac Island twice, with an AI as my coach both times: ChatGPT the first year, Claude the second. The Wall Street Journal ran a story this week about “datamaxxers,” people feeding every health metric they own into AI systems and building custom dashboards out of them. I never built a dashboard. But by this summer I was handing my coach heart rate files, glucose readings, sleep numbers, and questions about supplements. I became a datamaxxer in my own way. I just skipped the dashboard and went straight for the coaching.
What I Was Actually After
It was December 2024, and it had been a long time since my days of triathlons and open water swims.
What I wanted was a personal goal. Something that played to my strengths. Something I could put real time into that would make me healthier and happier. And something that was for me, not for anyone else.
I saw a post about the swim around Mackinac Island, the one Jon Voss organizes every year, and it looked like a decent challenge. Heather, my wife, could ride the island on her bike while I was in the water, so we would get the day together.
I was a college swimmer once. I did triathlons, including Ironman Lake Placid in 2008. The experience was in there somewhere. I just hadn’t done anything consistently in a very long time.
Year one, the goal was simple. Finish. Could I get around the island at all?
I did. August 10, 2025, in 4:53:39.
Two Seasons, Two Coaches
Year one, my coach was ChatGPT, and the early days were messy.
It helped me develop workouts, and that was about it. It never got the yards right. Math seemed hard. It would make up weird stuff, and I learned to check everything it handed me. But as the season went on, I improved regardless of my challenges with ChatGPT. The plan was imperfect, the workouts needed editing, and it was still more structure than I would have built on my own. Partway through that season, I bought an Apple Watch Ultra to record my swims, and I discovered I could feed screenshots of my results back to ChatGPT and get feedback. That was the first taste of a real loop. Swim, show it the data, adjust.
When I decided to swim the island again this year, I moved to Claude Cowork. I chose it for two specific reasons. I wanted documents on everything I was doing, not advice scattered across chat threads. And I wanted a coach with a consistent memory, one that knew in July what we’d decided in May.
The difference was a different league. Claude sharpened my goals before we ever built the plan. It built the plan, then refined it as my season unfolded. It wrote every workout and read every result. I used HealthFit to generate a proper swim log for each session, so my coach was working from the file off my watch instead of a screenshot. I had it analyze data from a CGM I wore to learn how my diet affected my swims. It helped me pick supplements. It looked at my sleep and told me how it was affecting my performance in the water.
And the biggest change from year one: I spent far less time editing the plans, the workouts, the advice, the feedback. Most of it was simply right, and my job shifted from correcting my coach to being coached.
That’s the two-year arc in one sentence. Year one, I supervised a tool. Year two, I worked with a coach.
What Claude Saw That I Couldn’t
One of the first things I did with Claude this spring was hand over the .fit file from the 2025 race and ask what had happened out there.
What came back was the story of my own race, and I hadn’t known any of it.
The story was in my heart rate data. Heart rate is the number that matters in an endurance event, because a five-hour swim is an exercise in managing your energy output, and heart rate is the honest measure of how fast you’re spending it. Mine showed I faded every single quarter. Q1 averaged 137.8 beats per minute. Q2, 133. Q3, 129.8. Q4, 128.5. My final ten minutes averaged 132 with no surge available, because there was nothing left to surge with. I also never went above 148 across nearly five hours, which meant I hadn’t blown up. I had opened too hot and then paid for it for four more hours.
Then it ran the counterfactual. If I had held Q1 at 130 instead of 137.8, that race likely ends 10 to 18 minutes earlier with no additional fitness at all.
That is the strongest case I can make for any of this. From inside the water, none of it is visible. You feel strong, then you feel tired, and you tell yourself a story about why. The data told a better story than I did.
It got compressed into one sentence I carried for a year. Q1 at heart rate 130. The thirty minutes lives in the first 75 minutes, not the last.
The Accountability Partner
This is the part I didn’t expect and the part I value most.
Every swim, I share the file. Every swim, I get back a debrief with the numbers and an honest read. Over months, that turns into patterns neither of us would have caught in a single session.
Here’s one example. Starting the first repeat of every set too fast is a habit I have had since college. It got named, and then it got counted. By the end of July, I had 14 consecutive sets where the first repeat came in under control. That is a small thing. It is also the exact discipline that cost me 10 to 18 minutes in 2025.
You can watch the work add up, too. Between December and May, I swam 59 times and covered 283,025 yards. Same week year over year, I was swimming 25% longer, eight seconds per hundred faster, at a heart rate 14 beats lower.
I have had human coaches. None of them were available at 5:30 in the morning to look at yesterday’s numbers with me.
Where It Fell Down
At the end of July, I noticed my Apple Watch was struggling to capture my heart rate in open water swims consistently. With Claude’s help, we figured out why. Cold water was knocking out the optical sensor, and my sleeveless wetsuit left bare arms sitting in that cold water. So I asked for ideas on how to fix it.
AI will optimize whatever you point it at.
I got back a lot of solutions. A chest strap to research, a two-device comparison protocol, more gear, more testing, more decisions. And I realized all of it was making things more complicated for me, with not enough time left to practice any of it before the swim in August. Mentally, that was creating stress instead of confidence.
So I told it to back off. I said this was getting too complicated, that I am trying to have fun while swimming, and that it isn’t a race. It agreed right away and wrote itself a note to default to subtraction from then until race day.
That exchange is the whole thing.
An AI coach will optimize whatever you point it at, as hard as you let it. It has no opinion about why you’re doing this. It never asked me whether all the measuring was making the swim better or worse, because that question was never in the data. I went looking for something that would make me healthier and happier. A system that is very good at making me faster will quietly trade one for the other if nobody stops it.
I had to be the one to stop it. And I had to do it more than once.
Here are two more caveats.
Two other things are worth knowing before you trust it too far.
It believes your sensors. My watch gave bad heart rate data twice, once from a loose band and once from the cold-water problem above. On one of those days I eased off my pace in response to a number that wasn’t real. Right instinct, wrong signal. The coach had no way to catch it, because it only reads what the file says.
It will also propose things that can’t work. When we were trying to solve my pacing discipline, it suggested a heart rate alert on my watch. Reasonable idea, except Apple doesn’t offer heart rate alerts in Open Water Swim mode. We spent a couple of weeks circling that before finding out, and by then the problem had solved itself through practice anyway.
Fifty-nine degrees
On July 25 the plan called for seven miles in open water. I got to Lake Michigan and found that an upwelling had dropped the water to 59 degrees, with big swells and a strong current. I was in a sleeveless suit. Two miles in I got sick, got out, and walked two miles back.
Jon had warned us the water was running cold this year, and that it can differ by several degrees on opposite sides of the island during the same swim. That came from someone who has watched this course for years, not from a model.
I was upset with myself. I had a plan for seven miles and I got out at two. Walking those two miles back, I felt like I had let myself down.
The good news is what happened next. My coach stopped me. It told me getting out was the right call, and it explained why: the body overrides the plan, every time. It logged the exit as a correct decision, left my readiness unchanged, and specifically did not try to make up the miles. Then it did something I didn’t expect. It helped me build a plan for that exact situation, so if those conditions showed up in August I would know what to do instead of hoping they wouldn’t.
That reset my expectations and gave me back some control over race day. The coach didn’t leave me sitting in a disappointed state of mind, which is where I was headed on my own.
It handled the aftermath better than I did. It just could not have made the decision.
AI is very good at the part before and the part after. The part during is still yours.
The 2026 Mack Swim
I went 4:47 over 8.64 miles.
My goal was 4:23:39, a thirty-minute improvement, and I missed it by a wide margin. Then I did the pace math.
Last year I swam 1:58.5 per hundred yards. Sunday I swam 1:53.2. Five seconds per hundred faster, over 334 more yards, in water that started at 69 degrees and dropped into the low 60s on the far side of the island. Jon was right. The first two hours were chop. The shadow side was glass. I fed every 700 to 1,000 arm strokes, and that worked well. I sighted on a point out on the horizon and drifted into deeper water instead of hugging the shore, which is how an 8.2-mile course becomes 8.64.
The clock says I fell short. The data says I got meaningfully faster in worse conditions over a longer line. Both are true, and I only know the second one because I have spent two years learning how to ask.
The AI Protocol, Mid-swim
Heather rode the island and kept up with me the whole way. She took pictures and put together a video of the day. She told me afterward that I looked grey and ashy climbing out.
I stopped once at British Landing, at mile 3.75, to see her and have her check the neck of my wetsuit. Chafing has bitten me before.
I have a written protocol for that neck. Hydrocolloid dressings, applied at five in the morning on clean, dry skin, no lotion anywhere near the adhesive. It is a good protocol. It came from the AI, and it worked.
The check was Heather’s hands.
How to Have Your Own AI Coach
You don’t need a system. You need a goal and a way to hand over real data. This is the order I’d do it in.
Pick the goal for yourself first.
Not a metric. A thing you want to be able to do, that plays to something you’re already decent at. AI is very good at building a path to a goal and completely uninterested in whether it’s the right goal for you.
Give it the raw file, not your memory of the workout.
This is the single biggest change. “I felt strong for the first hour” is a story I tell myself. The .fit file off my watch is what happened. Every real insight I got came from that file and not from my description of the swim.
Here is my entire setup, and it took about 10 minutes to put together. I swim with an Apple Watch. On my iPhone, I run an app called HealthFit, which reads the workout out of Apple Health and writes it as a .fit file, the standard format most serious training tools use. I have it set to export automatically, so the swim is waiting for me before I’m dried off. Several sites walk through that configuration step by step, so I won’t repeat it here.
Both iCloud and Google Drive are connected to my Mac. Grabbing the exported file and dropping it into the Google Drive folder where I keep my swim history takes a few seconds. Then I tell Claude the file is there and ask what happened.
That’s the whole thing. A watch, one app, a folder, and a question. No dashboard, no integration, nothing to maintain.
The file holds more than you would expect. Every length and every interval. Heart rate second by second. Stroke count, rest between repeats, the exact minute I sped up when I shouldn’t have. That is the difference between a coach guessing and a coach reading.
Make it write things down.
Ask for the plan as a document you can both come back to and edit. Ask it to keep a running memory of what it has learned about you. Chat threads disappear. Documents compound.
Debrief every session, not just the important ones.
Mine takes five minutes. The numbers, an honest read, one thing to carry into the next swim. The value isn’t in any single debrief. It’s in month four, when it can tell you what you’ve been doing for eleven weeks straight.
Have it name one habit and count it.
One. Mine was opening the first repeat too fast. Naming it made it visible and counting it made it change.
Say out loud what you’re actually after, and then say it again.
It will drift toward optimizing, because that’s what it’s good at. You have to keep telling it what winning means to you.
Keep the in-the-moment decisions.
Whether to get in, whether to get out, whether today is the day. Nothing in the plan knows what the water is doing.
The Future
So, is AI ready to be your full-time coach? For the training, way closer than I expected. It wrote every yard I swam this year, and it reads my races better than I can read them myself. For the rest, no. It can’t want this for me, it can’t feel 59-degree water, and it will never be standing at British Landing. Those were always the human parts. What I didn’t know was how much of the rest could come off my plate, and how much better I would get once it did.
Next year I plan to swim across the Straits. The AI tools will continue to improve, and I’m excited to see what that will be like. There will also be new wearables and supplements I can add into my plan. This is a different challenge than going around the island. I know I can do the distance. What I don’t know is what the experience will be like, and nothing in a training plan is going to tell me that before I get in the water.
That’s why I picked this in the first place.