What should you ask an AI about your Garmin data?
Almost anything, once you have told it what the numbers actually mean, and almost nothing before that. The prompts below are not "10 fun questions for your AI." They are corrections: each one exists because we tested a Garmin metric against real outcomes and found a gap between what the number seems to say and what it measures, and an unbriefed AI falls straight into that gap.
This guide assumes your data is already connected; if not, the setup guide covers all four routes. Everything below works the same in ChatGPT, Claude, or any assistant that can see your training data.
Why the briefing matters more than the questions
A language model treats confident inputs as true. Garmin's metrics are nothing if not confident: a race prediction to the second, a readiness score to the point. We tested two of the most prominent numbers against real outcomes, and neither means what it appears to mean:
- Across 22,422 daily prediction snapshots, the race predictor sat a median 20 to 23% faster than everyday training bests — yet within 0.6% of actual race-day results in the races we could check. It is a race-day ceiling, not a training pace.
- Across 5,136 paired steady runs, morning Training Readiness correlated 0.056 with how the day's run actually went, while tracking yesterday's training load at −0.272. It is load accounting, not a forecast.
Cohort caveat, as always: self-selected users of our former analytics platform, likely fitter and more data-curious than average.
An AI that takes both numbers at face value will set your tempo pace off a race-day ceiling and cancel workouts because an accounting metric looked low. So the first thing you give it is not a question. It is a briefing.
The briefing paragraph
Paste this once at the start of a training conversation, or save it wherever your assistant keeps standing instructions:
When you read my Garmin data, apply these rules. Treat the race predictor as a race-day ceiling, not a training pace; training paces should come from recent actual sessions. Treat Training Readiness and recovery time as a record of recent load, not a forecast of today's performance. Treat Body Battery, HRV status and sleep score as one recovery signal, not three independent confirmations — they are built from overlapping inputs. Treat the VO2 max value as a slow trend line; ignore day-to-day changes. Weight how I say I feel over any score, and say so explicitly when a score and my stated feel disagree.
Every rule in that paragraph traces to a study in our research library. That is the whole trick. The rest of this guide is what to ask once the AI is calibrated.
Prompts for the weekly review
The weekly review is where a connected AI earns its keep, because pattern-across-sessions is exactly what a chat interface does better than a dashboard.
Summarise my last 14 days of training. Separate what I did (volume, intensity distribution, load) from how I responded (resting HR, HRV trend, how runs at a given pace felt by heart rate). Flag anywhere the two diverge.
Compare my average pace at heart rates between 140 and 150 this month against last month, same conditions where possible. Is there evidence I am getting fitter, or just evidence I am training more?
Which of my last ten hard sessions was followed by the worst two days of recovery markers? What did that session look like?
The divergence framing matters. "What I did" comes from the watch's strong suit (measurement); "how I responded" is where its composite scores get shaky, so asking for the raw markers (resting HR, HRV trend, pace-at-HR) keeps the AI on measured ground.
Prompts for race planning
My race predictor says X for a 10K. Based on my actual training bests over the last eight weeks, not the predictor, what pace range have I demonstrated I can hold, and what would a realistic race plan look like?
I race in three weeks. Build a taper from my current weekly load, and tell me what my load numbers should look like each week so I can check against the watch.
The first prompt exists because of the 20–23% gap. If you ask "can I hit my predicted time," the AI reads the prediction as evidence. If you ask what your training has demonstrated, it reads your sessions instead, and the sessions are the part the watch measures well.
Prompts for the readiness argument
The most common daily question is some form of "the watch says I'm not recovered, but I feel fine." Our data says the score is tracking your recent load, so make the AI arbitrate with that in mind:
My Training Readiness is low this morning but I feel good. Look at my last three days of load: is the score low because of what I did, or because of how I slept? If it is load accounting, tell me what an easy version of today's planned session looks like. If my sleep and HRV genuinely cratered, say so.
My readiness has been high all week. Before I stack another hard session: what does my acute-to-chronic load ratio look like, and is there anything in my recent data the score would not weight — a race effort, travel, unusually hot sessions?
The second prompt is the guard rail for the opposite failure: top-readiness days produced no better runs than middling ones in our data, so a green score is not a license.
What not to ask
An AI on Garmin data inherits the data's blind spots, and no amount of prompting fixes these:
- "Am I injured / getting injured?" Nothing in the data measures tissue tolerance. Load spikes are a risk marker, not a diagnosis, and the AI will produce confident-sounding speculation if pushed. Ask "has my load ramped faster than my usual pattern" instead, and take body-part-specific pain to a human.
- "What does my sleep score say about tonight?" In our data, last night's sleep score correlated 0.04 with the next day's run. Chronic sleep debt matters; the nightly score barely predicts anything by itself.
- Anything during a workout. Every connector route syncs after the fact, minutes to hours behind. The AI is analysing your training history, never your current interval.
Frequently asked questions
Do these prompts work the same in ChatGPT and Claude?
Yes. The prompts are plain language and the briefing paragraph is plain language; nothing here depends on one vendor's features. What differs is where you store standing instructions — a Claude project's instructions or ChatGPT's custom instructions both hold the briefing fine.
Why not just let the AI read the metrics as Garmin intends them?
Because two of the headline numbers, read as intended, point the wrong way: a race predictor that reads as a training target and a readiness score that reads as a forecast. We tested both against outcomes — 22,422 prediction snapshots, 5,136 runs — and the briefing encodes what the tests found. An unbriefed AI is not neutral; it inherits the marketing.
Will this make the AI disagree with my watch?
Sometimes, and that is the point. The useful output is not "the watch is wrong," it is the explicit arbitration: this score is low because of what you did yesterday, your raw markers look normal, here is the easy version of today. You stay the judge.
Does the briefing apply to other brands' scores?
The specific numbers are Garmin findings. The structure travels: recovery composites from every major brand are unpublished formulas over the same short input list, so "one signal, not three confirmations" and "weight stated feel over scores" are safe defaults anywhere. Our cross-brand recovery comparison goes deeper.
Steal the paragraph
If you take one thing from this page, take the briefing paragraph. It is the difference between an AI that amplifies the watch's confidence and one that corrects for it. And if your data is not connected yet, start there; the prompts are worthless without something to read. The rest of our Garmin and AI collection covers the connectors, the coaching tools, and what each one can actually see.
