Why ChatGPT training plans miss, and what fixes them

The structure of a ChatGPT running plan is usually fine. Easy runs, one long run, one quality session, a cutback week every so often: it's the shape every coaching book agrees on. Where these plans go wrong is the number on each line: the pace. And the pace comes from whatever you told ChatGPT about your fitness, which is usually a guess, or worse, your watch's VO2 max.

The VO2 max trap

Here is my own case. On October 9, my Apple Watch put my VO2 max at 61.3. Five days earlier I ran an official 5K in 22:20.

The equivalences come from Jack Daniels' VDOT model, the one most running calculators use. A plan built from the left column asks for intervals a full minute per kilometre faster than I can run them. The first session fails, and the conclusion you draw, that you're not fit enough for the plan, is the wrong one. The plan wasn't built for you.

Apple's VO2 max is an estimate from outdoor walks and runs, and it is most useful as a trend: mine rising from 44 in May to 61 in October says something true about the last five months. As an input for paces, it's the wrong number. A race result, even a parkrun, is a measurement.

The fix: let it read what you ran

ChatGPT and Claude can read your Apple Watch workouts directly once they're connected to them. Then the question changes from "here's my VO2 max, write me a plan" to "you can see my training, set my paces". This is what came back:

The answer did the one thing a blind plan can't: it looked for the best evidence available, found a race, and preferred it to the watch's estimate. As a sanity check, my running app's targets for interval reps that same week sat between 4:10 and 4:30 per kilometre: right where the race says, nowhere near 3:19.

Three more reasons plans miss

• No starting volume. "Run 40 km next week" is fine if you ran 35 last week and dangerous if you ran 15. An assistant that can read your last month can start from it.
• Heat. A summer easy pace is slower at the same effort. Plans written in paces and not in heart rate get harder every hot afternoon.
• No week two. A plan written once never hears how week one went. The real value is in asking again each week, with the sessions in front of it.

A prompt that works

Once your workouts are connected, something like this gets a plan built on you rather than on an average runner:

"Read my last six weeks of running. Set my training paces from my most recent race, not from my watch's VO2 max, and tell me which race you used. Then build next week: four runs, one quality session, and the total volume no more than ten percent above my average week."

Asking it to name the race it used matters: it forces the answer to show its evidence, and it tells you when there isn't any. For more questions worth asking, see five prompts for an AI running coach.