5 prompts for an AI running coach, with the real answers

Most "AI running coach prompts" you'll find are role-play: "act as an elite running coach and write me a 12-week plan". They produce a nice-looking plan, built on nothing. The prompts below are different: they're questions about what you actually ran, and they only work if the assistant can read your workouts. I asked each of them to Claude, connected to my Apple Watch data through HealthSync. The answers are the real ones.

1. Compare two sessions of the same kind

The fastest way to see progress is two sessions with the same structure, a few weeks apart. Name both, or describe them, and let the assistant find them.

Compare my two 400 m sessions.

healthsync · get_workout 2 workouts · 11 reps of 400 m

Fifteen days apart, about 13 seconds faster per rep. On September 15 you ran six 400s between 1:57 and 2:02; on September 30, five between 1:42 and 1:49. Heart rate peaked a little higher, 151 to 155 against 150, for a much bigger gain in speed.

Why it works: it compares like with like. Average pace across all your runs mixes easy days and hard days; the same session twice doesn't.

2. Ask how a race was paced

Your watch tells you your time. It doesn't tell you whether you ran it well.

Did I pace my 5K well?

healthsync · get_workout_series distanceWalkingRunning · heartRate · 4 laps

Yes, a clean negative split. 4:36 for the first kilometre, 4:32 over the next two, 4:31 for the fourth and 4:24 for the last 940 m. Heart rate settled at 157 to 159 through the middle, sat around 160 in the fourth kilometre and only peaked, at 165, in the last. There was room to go out a few seconds faster.

Why it works: pacing is in the splits and the heart rate together, which no summary screen shows side by side. The last sentence is the useful one: it tells you what to try next time.

3. Ask for a record, with a date range

"What's my longest run" is ambiguous: since when? Give it the window you care about.

What's my longest run since I started training in May?

healthsync · list_workouts running · 2026-05-01 → 2026-10-09

12.03 km, on October 7, in 1:14:45 at 127 bpm. It only just beats September 23, 12.01 km, which was a progression run rather than an easy one. Before that, nothing past 10.1 km.

Why it works: the date range turns a trivia question into a training one. And the assistant noticed on its own that the two longest runs weren't the same kind of run.

4. Ask about a trend in your recovery

The numbers outside workouts move slowly and say a lot. Ask for the trend, not today's value.

How has my resting heart rate changed since May?

healthsync · list_health_metrics restingHeartRateBpm · weekly · 2026-05-04 → 2026-10-09

Down 17 bpm. It averaged 60.3 bpm the week of May 4 and 43.4 the week of September 28, its lowest. The drop was steepest from May to mid-June, 60 to 49 bpm in five weeks. Then it slowed: 46 to 50 bpm through July, 46 to 49 from August to mid-September, 43 to 45 since.

Why it works: a weekly trend smooths out the noise of single days. Resting heart rate falling while training goes up is the most reliable sign that the training is working.

5. Ask for a season total, and say what to count

Totals are where duplicate recordings bite: a ride saved by your watch and by a bike computer counts twice in a careless sum.

How much did I ride this summer?

healthsync · list_workouts cycling · 2026-06-01 → 2026-08-31 · 42 rides

989 km over 42 rides, 55 h 42 in the saddle and 6,237 m of climbing, counting what your Apple Watch recorded: the rides your Garmin also saved aren't counted twice, and the few only the Garmin recorded aren't counted at all. Six of them went past 40 km, the longest 90.5 km on July 1.

Why it works: the answer said which recordings it counted and what that leaves out. If you record with two devices, ask it to.

What makes these prompts work

A date or a range. "Since May", "this summer", "September 15".
A comparison. Two sessions, two months, before and after. A number on its own means little.
A request for the numbers. Ask it to quote what it used. If it can't, it's guessing.
A question you'd ask a coach, not an instruction to play one.

For the bigger question behind all of these, see can AI be your running coach?

5 prompts for an AI running coach, and what they really return