How to ask ChatGPT about your Garmin data.

What makes a question worth asking

Your Garmin app is very good at showing you a number. It's the fastest thing in the world at telling you last night's sleep total. If that's what you want, open the app — an AI round trip is strictly slower.

What the app can't do is hold five metrics across two weeks in view at once and notice that two of them moved together. That's the gap you're filling, and it tells you exactly what shape your questions should be.

1. It spans time

A single reading has no meaning without the readings around it. An HRV of 38ms is neither good nor bad — it's alarming if your average is 60 and unremarkable if it's 40. Every useful question has a window in it, explicit or implied: this week versus my baseline, the last 14 days, since I got back.

2. It needs more than one metric

If one number answers it, the app already did. The value appears when the answer lives in the relationship — sleep against training load, HRV against resting heart rate, overnight recovery against time in bed. That's a join across sources, and it's tedious enough by hand that nobody does it.

3. It can come back "nothing's wrong"

This one is a filter on you, not the model. If the only answer you'd accept is "yes, something's up," you're not asking a question — you're asking for agreement, and a language model will cheerfully give you that. A real question is falsifiable. "Is my load outpacing my recovery?" can come back no, and the no is worth as much as the yes.

4. It asks what the data shows, not what to do

"Should I race on Sunday?" isn't answerable from HRV. The model will answer anyway, fluently, and you'll have outsourced a decision to something that has no idea how much the race matters to you. Ask "what does my recovery trend say heading into Sunday?" and you keep the decision, which is where it belongs.

A question that actually worked

A real one, against a real Garmin account — one of our users, shared with their permission:

Look at what that question is doing. It spans time (Sunday, against the days around it). It needs several metrics — Body Battery alone can't explain Body Battery, so answering it means pulling sleep, recovery, and workouts together. It's falsifiable: the honest answer could easily have been "you did a hard session Saturday, this is normal." And it asks what happened, not what to do about it.

The answer found no workout at all that weekend, so training wasn't the cause. It traced to a single night: Garmin logged 6h40m with REM of 10 minutes, against 102, 89 and 84 the three nights before, and the next morning showed resting HR 63 against a 51.5 average with HRV at 38. Then the part that mattered most — full rebound inside 24 hours, reframing it from "something's wrong" to "one bad night, handled."

But the follow-up is where it got interesting, and it's a good lesson in not stopping at the first answer. Asking the same question of a second device — a Whoop worn the same night — produced a flat contradiction. Whoop logged 10h35m and 259 minutes of REM for the night Garmin scored at 6h40m and 10 minutes. A 25× gap, same body.

Neither device is broken. Bedtime was around 02:00; Garmin closed the sleep session at 09:07, while Whoop kept counting to 14:15 and caught the sleep that came after. The two devices agreed almost exactly on the recovery — Whoop 94 → 32 → 99, Garmin Body Battery 72 → 25 → 94 — and disagreed entirely on the sleep behind it, because they weren't measuring the same night. The full conversation is on the Garmin page.

The lesson generalises past this one night: a single device's number is a measurement, not a fact. If the answer matters, ask whether anything else you own saw the same thing — and expect the absolute values to differ even when the story doesn't.

Notice too what the model refused to do: name a cause. That pattern is consistent with a late night, a drink, a big meal, or something viral starting, and the data cannot distinguish between them — nor can it tell you how much of anything was involved. A good answer stops there. If yours doesn't, be suspicious of it.

Questions that work

• "Is my Body Battery actually recovering overnight, or just topping out lower each week?" — a trend across daily peaks, not a reading.
• "My HRV status went Unbalanced. What changed in the two weeks before that?" — forces a correlation across load, sleep and resting HR.
• "Has my training load climbed faster than my sleep can absorb?" — two series, compared.
• "Compare my lap splits from this week's intervals to the same session in March." — two workouts, aligned; genuinely tedious by hand.
• "Did my VO2max estimate actually improve, or is that noise?" — asks for the variance, not the number. Good questions often ask about the error bars.
• "Where does my Garmin disagree with my Oura about last night?" — two devices at once, which is the one thing a single-device connector structurally can't do.

Questions that waste the round trip

• "What was my sleep last night?" — one metric, one moment. The app is faster and you already have it open.
• "How many steps today?" — likewise. You're adding latency to a glance.
• "Am I overtraining?" — too vague to be falsifiable. You'll get a confident essay assembled around whichever numbers looked most dramatic. Name the window and the metrics and it becomes a real question.
• "Should I race on Sunday?" — not in the data. A good model will tell you that; a bad one won't.
• "Am I healthy?" — not a question a wearable can answer, and not one to outsource. These are consumer sensors, not diagnostics.

Two things worth knowing

The model interprets; the data is just data

ghurt supplies numbers with their context — units, baselines, date ranges — so the model reasons about your actual values rather than guessing. The interpretation is still the model's, and models are confident even when wrong. If an answer surprises you, ask which metric drove the conclusion. It'll show you, and sometimes you'll find the reasoning was thinner than the tone suggested.

Your data passes through your AI provider

ghurt doesn't store your health data — it's fetched per question and passed through. But "passed through" means it reaches OpenAI or Anthropic as part of your conversation, under their privacy policy, not ours. That's inherent to asking an AI about your body, whichever connector you use. Whether it's an acceptable trade is genuinely your call, and we'd rather you made it knowingly.

If it stops working

Garmin changes its session mechanism without notice, and when it does, everything built on it breaks at once — that's the normal weather here, and it's why DIY scripts are a recurring chore rather than a one-time build. In March 2026 it took out garth, the library most of the ecosystem logged in through.

The difference with a maintained connector is whose problem it is. When Garmin changes something, fixing it is our job — you shouldn't need to do anything. If your connection does break, email us.