A week with an AI running coach

Sunday evening, just after nine. Your phone buzzes: next week is in the calendar. Four sessions, threshold on Tuesday, the long run on Saturday instead of Sunday because you said Sunday is taken, and a line on why the threshold session is shorter than last week's.

That is how every week with Coach Fartlek starts. It is our own app, so expect bias. What the coach does between sessions is harder to explain than what it does in the chat, so here is an ordinary week, with the hours and thresholds that are actually in the code.

Sunday evening: the week is laid out

At nine in the evening on the last day of your training week, next week gets written. Three hours earlier a short message can arrive, but only if something looks off: you have no free time slots entered, or you mentioned a work trip in the chat that the calendar does not show. If the week looks normal, nothing arrives.

The input is last week as it turned out, not as it was meant to be. Which sessions got done, how the easy runs were actually run, HRV and resting heart rate over the last two weeks, which hours you have free, whether you have a niggle logged in the health log, and the rules the coach has learned about you over time. "Does not tolerate two quality days back to back" is one such rule, saved after it went wrong twice, and it comes along into every new week.

The plan is a hypothesis. That is the exact wording in the coach's own principles, and it is also how the plan is treated for the rest of the week.

Tuesday morning: the reminder

At eight a message about today's session arrives. If the session has a start time in the calendar it arrives four hours before, and if the session starts before ten in the morning no reminder comes at all, because we send nothing before six. It is two to four sentences on what the session is for, with the weather at the hour you will be running woven in. If you slept short before a hard session, the coach proposes the softer version and says why in a clause, fewer reps or easy pace instead, without moralising about sleep.

Sometimes it is just a reminder. An easy run on an ordinary Tuesday needs nothing more than that.

Tuesday evening: the session is read before you are out of the shower

The watch syncs to Intervals.icu, which passes the session on, and a few minutes later the analysis is in the chat. The session is judged against what it was meant to do, not against the name of the session type. If the intent was four times eight minutes at threshold, the question is whether you held threshold pace with sensible rest, not whether the total distance matched.

If you have run the same loop before, the coach compares against it as is, since the terrain is then identical. If today was warmer, that explains a higher heart rate at the same pace, and it says so. You are spared the stat dump: zone percentages and heart-rate drift are input to the verdict, not findings to be listed.

The message ends with at most one question, and only if the session gave a reason for one. If you write how it felt, that is weighed against the numbers, and a gap between the two is usually the interesting part. If you write nothing, the conclusion stays conditional, because the data shows what happened and never how it felt.

If the session changed what should come next, an aborted run or a new pain, the coming days are adjusted in the same pass. Then you read what it means for you, "we keep Thursday easy", and nothing about the plan having been reviewed.

Wednesday: the day you mostly hear nothing

Every morning at nine a check runs that usually ends in silence. It looks for a handful of signals: HRV clearly below your normal range last night, resting heart rate at least three beats above your normal ceiling, a heat forecast that makes today's pace targets unrealistic, a niggle in the health log you have been quiet about for two days, or a session yesterday that did not happen. If none of them is there, nothing happens.

On Wednesday afternoon the coach also reads your recovery for the week so far, against two weeks of HRV and resting heart rate. If it looks like an ordinary week, the assessment is stored as a note and not sent. Only what stands out reaches you, in either direction.

If you want an app that checks in every day, you will find this one quiet.

Thursday: the morning HRV drops

Now an example. HRV is below your range for the second morning in a row, resting heart rate has crept up, and Friday has intervals on it. At nine a message arrives that says which number triggered it and what changed: the intervals are moved to Saturday and Friday is an easy run. The change is already made in the calendar, and it is as small as possible within the week.

What the coach does not know is why. A low HRV looks the same after a glass of wine, an oncoming cold, a night with a sick child and a build that has started to bite. If you write "slept four hours, the kid was up", the answer changes. If you write nothing, the coach assumes training load, and that can be wrong.

A value just under the range moves nothing, because a dip the day after a hard session is expected. Only when the morning sits clearly outside your own range, or when several in a row point the same way, does something change.

Saturday: the week falls apart

The long run was supposed to be Saturday. It did not happen. You write that in the chat in your own words, and the coach rewrites the rest of the week accordingly: which session is saved and which is dropped. If the change is big enough, a race that moves or a week that has to be rebuilt from scratch, a heavier planning pass takes over in the same reply.

Sunday evening at nine it starts again.

What you do not get

The app needs a watch connected through Intervals.icu, which pulls from Garmin, Coros, Polar, Suunto and most others. Without a watch you can log sessions by hand, but most of the week above disappears with it. The coach covers 5K to the marathon, not ultra and not multisport.

The coach is a language model. It reasons well about training theory and can still be wrong in the same calm tone as when it is right. Ask for the reasoning when something feels off, and expect your own setup to sometimes be better than the coach's objection. Can AI give useful input on your running describes that situation in more depth, and for anyone who wants to plan entirely on their own there is a mode where the coach only reviews.

The week above assumes Pro or Elite. The free tier is a trial with a fixed number of messages that does not refill, and the coach's own messages draw from the same pool, so it runs out in a few days if the coach is allowed to lead. What the tiers cost is on the pricing page.

If you have already tried letting ChatGPT run your training, you know where it ends: the sessions live in the watch and never reach the chat, and nothing happens until you ask. Why ChatGPT forgets your training goes through it. How the recovery part works in detail is in the running app that adapts to your recovery.