Your AI running coach, with live Garmin or Strava data.

You wake up, glance at your watch, and wonder if today is the threshold day or the easy day. Instead of staring at one composite "readiness" score, ask the AI to weigh the actual signals.

Your AI calls four tools: overnight HRV, sleep score, body battery, and yesterday's training load. It cross-references each against your seven-day baseline and returns something like: "You're recovering, but not back to baseline. HRV is 9% below baseline, body battery topped at 67. Best move: 60 to 70 minutes Zone 2, cap HR at 145."

That's the answer a coach would give if they had your numbers in front of them. Faster than reading three dashboards yourself, and it remembers the context for the next question.

Sunday evening reflection. You want a real read on the week before deciding what next week looks like. Open Claude or ChatGPT, ask one question.

The AI pulls every activity, plots training load by day, and reads polarization. It tells you whether you nailed the easy/hard ratio, where the load actually came from, and which session defined the week. It catches things that are easy to miss, like a single threshold session being 48% of weekly load, and surfaces them with the math.

Output: stat tiles (load, volume, elevation, active time), a bar chart by day colored by activity type, and a paragraph or two of analysis. The same view a good coach would draw on a notepad.

You've been training for a 10K. You want to know if your goal is realistic and what you need to hit per kilometer.

The AI checks Garmin's race predictions (built from VO2 max and recent run history), compares them to your stated goal, and works out the gap. Sample answer: "Current 10K prediction: 46:12. To break 45 you need an average pace of 4:29/km. Your last 5K time trial was 21:47, which projects a faster 10K than the predictor shows. The aerobic base is there. The limiter is threshold pace." Plus suggested sessions to close the gap.

Same loop works for 5K, half marathon, marathon. Ask "Build me an AI marathon plan from current fitness" and the conversation continues into a four-week block tailored to where you are.

The hardest part of long runs is reading them after. You think it went okay; the watch isn't sure. Ask the AI to look at the splits and the heart-rate curve together.

It pulls activity details, splits, and HR zones. Computes HR drift across thirds of the run (a key marker of fueling and pacing), notes which kilometers were honest Zone 2 vs. drifting up to threshold, and reports on aerobic training effect. Sample: "HR drift was 4.2% from km 12 to 22. Above 5% suggests fueling or pacing issues, you're in the safe band. Aerobic TE 3.8 means a productive long run that didn't dig into anaerobic reserves. The 1km surge at km 18 was where pace fell off, HR had already crept above 162 by then."

Your watch reports a sleep score every morning. What you actually want to know: are the patterns lining up the way they should? Are heavy sessions impacting recovery enough that you should change something?

The AI cross-references training load with each night's sleep stages. It looks for the pattern athletes know but rarely quantify: sympathetic nervous system activation from threshold and intervals reducing REM that night. Sample: "Yes, three of your last four hard sessions show REM under 1h the same night, vs. your baseline of 1.6h. Your weekday easy runs don't show the pattern. The interval session on April 30 had the biggest hit."

From there you can ask "What if I move intervals to the morning?" or "Does deep sleep recover?" and get a real follow-up.

Three hard weeks in. Legs feel heavy, motivation dipping. Useful to read multiple recovery signals together rather than fixating on one.

The AI looks at HRV trend over 28 days, resting heart rate trend, body battery accumulation overnight, and stress score during the day. It reads the constellation. Sample: "HRV is 12% below your 28-day baseline, RHR is up 4 bpm, body battery only recovers to 78 overnight. Three signals pointing the same way. This isn't deep-in-the-work fatigue, it's accumulating. Recommendation: take a deload week, drop volume by 30%, no intensity."

You can challenge it: "What if I keep volume but cut intensity?" The AI will recompute the load and tell you whether that gets the system back.

You've been doing threshold sessions on the bike. You want to see how each one compared, not just by average power but by how steady you were and whether the legs held up across the ride.

The AI pulls activity details and per-second streams for each ride. Computes normalized power vs. average power (proxy for effort variability), and looks at HR/power decoupling across the ride. Sample: "Threshold session 3 had the cleanest power profile: 245W average, 252W NP (1.03 ratio = very steady), 1.9% decoupling. Session 4 was harder on paper (258W NP) but decoupling was 6.4%, which means you were riding above sustainable. Session 3 was the better executed workout."

Strava users get this in full depth, since per-second streams are a Strava strength.