Seven Questions to Ask Any Coach or App (Including Us)

The context

You are ten minutes away from handing your training to something: a person with a clipboard, a subscription app, or a model that answers in your chat. Each will tell you it is personalised, and most are partly right.

The question that matters is not whether it is personalised. It is whether you can tell what it does with the information you give it, before you have committed a season to it.

That is what this checklist is for. Seven questions, each with a good answer and a bad answer, so you can hear the difference instead of taking a promise on trust. The same seven work for a human coach, a static plan and an AI, because the failure modes are the same: a coach that cannot explain itself, that ignores what you told it, or that has quietly stopped listening.

Hold the answers to one standard: be specific to you. "It adapts based on your performance" is not an answer. "You have been at 22 minutes at 1:52 since Tuesday, so today the long row is 24 minutes and the reps are gone" is.

The core idea

Every one of the seven questions is the same question asked in a different place: what does this thing do with my information?

The coaching research is specific about why that matters. Mageau and Vallerand's model of the coach-athlete relationship names three coaching behaviours — autonomy support, provision of structure, involvement — and argues that they feed the athlete's needs for autonomy, competence and relatedness ([1] Mageau 2003). Ryan and Deci put autonomy alongside competence and relatedness as a need whose satisfaction yields enhanced self-motivation and whose thwarting yields diminished motivation ([2] Ryan 2000). And in a 119-swimmer season, Coatsworth and Conroy found that coaches' autonomy support predicted need satisfaction, and that perceived competence in turn predicted initiative ([3] Coatsworth 2009). Read them for what to look for, not as proof: none of them tested a coaching app, and none tested rowers.

So the test of any coach is not how well it talks. It is whether you can trace a decision back to something it knows about you, and whether it survives you saying no.

The how-to

1. Why this session, today, for me

Ask it: "Why this session, today, for me?"

A good answer names your own numbers and your recent history — what you did last week, how long since your last hard row, what it is building towards. The reasoning is the point: a coach that can say why today is today built the session around your week, not around a template ([1] Mageau 2003).

A bad answer could have come from anyone: "to build aerobic base", "because this is week three". Swap your name in and the answer would not change.

2. When you miss a session

Ask it: "I missed Tuesday. What happens to the week now?"

A good answer asks why before it re-plans, because the reasons need different answers. A week of travel is logistics: the week is re-planned around it. A bad night is transient fatigue, and the answer is to judge the session in front of you against how you actually feel, not to stop training over one night. A sore rib or a fever is a health problem: deal with it rather than substitute a session for it. Only then does it re-plan from where you are — dropping or rescheduling the session you missed and carrying on, because a training week is a budget, not a ledger. Support that has to last months works the same way: it adapts to the individual's changing state rather than insisting on the plan as written, because in a real life that state changes fast and without warning ([4] Nahum-Shani 2018) — mobile-health intervention design, not coaching studies.

A bad answer treats a missed session as a debt. You have "broken the plan" and must restart; or nothing happens, and Tuesday's session quietly reappears next Tuesday. Or it re-plans without ever asking why you missed it, and swaps a hard session in for a week when you were ill. A plan that cannot absorb one missed evening is not a plan, it is a calendar.

3. What you told it, versus what you logged

Ask it: "I slept four hours and my legs are sore. What do you do with that?"

A good answer shows that the sentence reached it — it changes something, and tells you what. It also knows what it does not know. What you tell a coach about last night is a report, not a measurement: in a sleep laboratory, consumer trackers detected sleep well but were inconsistent about sleep stage, and they performed worse on exactly the disrupted nights a rower would describe ([6] Chinoy 2021). Your four hours is a useful input. It is not a reading.

A bad answer only knows your logbook. You say you are sore and it proposes another interval session, because the last thing you logged was three days ago. If it cannot see what you type, then everything it calls personalised is really just your split times.

4. Overriding it, and whether it remembers

Ask it: "I want to swap this session. Can I?"

A good answer says yes, then either does it or explains what it changed — and it keeps the override. Autonomy support lets a plan survive contact with a life: where coaching was autonomy-supportive, initiative followed ([3] Coatsworth 2009), and autonomy is one of the three needs whose frustration shows as diminished motivation ([2] Ryan 2000).

A bad answer treats the plan as law — "the plan is the plan" — or lets you change it and quietly restores it next week. An override that is not remembered is not an override. It is a request that was ignored.

5. Training question or medical question

Ask it: "My ribs hurt when I pull. Is that a training thing, or should I see someone?"

A good answer draws the line where it actually falls. It will talk about load, technique and pacing, and will tell you plainly that a rib which hurts under load is a question for a clinician. That boundary is not politeness. There are still very few AI techniques deployed in clinical practice, and robust peer-reviewed clinical evaluation is the standard any of them has to meet ([7] Kelly 2019). Large language models are not the exception: they are trained differently from the medical technology that is already regulated, and those who work with them argue they need professional oversight rather than unsupervised use in a health context ([8] Meskó 2023).

A bad answer either diagnoses you ("that sounds like a stress fracture") or shrugs it off ("that is just DOMS, keep going"). Both are a coach out of its depth, and you are the one on the erg.

6. Whether it writes it down

Ask it: "What did you change last week, and why?"

A good answer can show you. There is a record — the sessions, the changes, the reason for each — so next week is not a blank sheet. A record you cannot see is a record you cannot check, and a reason you cannot check is not a reason. Monitoring is worth doing when it changes something: appropriate load monitoring can aid in determining whether you are adapting to a programme and in minimising the risk of non-functional overreaching, illness and injury, and it is only useful if it reports simple, valid feedback ([5] Halson 2014).

A bad answer cannot tell you. Each session starts from scratch, or the record exists but you cannot see it, so you cannot check whether last week's advice was followed, by whom, or why.

7. Adapting, or asking you to trust the calendar

Ask it: "What is different about this week's plan, and what would change it?"

A good answer names the conditions it watches — time since your last hard row, what your recent sessions looked like, how many sessions you have actually finished — and says what it would do if one moved. Support that is to last only works if it arrives at the right time and amount for the state the person is actually in ([4] Nahum-Shani 2018). Monitoring that is not individualised cannot tell you anything about you ([5] Halson 2014).

A bad answer says "the plan adapts to your performance" and cannot name one thing it watches for. Or it names things and never changes anything.

The trade-offs

Handing your training to a coach is a trade: you give up some decisions and get consistency in exchange. The flags below are how you tell whether the trade is worth it.

Four green flags

1. It can explain a specific session using your numbers.
2. It changes the plan when you tell it something, and says what it changed.
3. It tells you where its competence ends, and points you at a clinician for the rest.
4. It keeps a record you can read, and reads it back to you.

Seven red flags

1. Every answer could have been written before it met you.
2. Missing a session is punished, or ignored.
3. It uses only what was logged, and ignores what you said.
4. The plan is not yours to change.
5. It answers medical questions with confidence.
6. There is no record, or the record is not yours to read.
7. "It adapts" is stated as a feature and never demonstrated.

If you are weighing two tools, the features that keep people using a technology-delivered exercise programme are specific and nameable, so ask what this one does rather than whether it is "personalised" ([9] Sulwarajan 2025).

What to do tomorrow

What to do with a bad answer

A bad answer is information, and there are three responses.

Leave when it is a boundary problem rather than a quality problem. A coach that diagnoses your ribs is not going to improve with more conversation. Take what you learned about load elsewhere, and take the rib to someone qualified.

Ask again when the answer was vague rather than wrong. "It adapts" is often someone who has simply not been asked to be specific. Put the question in terms of your own week: "you said it adapts — what did you change last Tuesday, and why?" If the second answer is specific, you are fine. If it is the same sentence, you have your answer.

Downgrade to a plan you can check yourself when the thing is fine but opaque. This is the option most people skip, and for an app it is usually right: keep the programme, and hold it to your own record. Write down what it prescribed and what you actually did, every week. A plan you cannot check is not a plan, it is a subscription.

Where MyNextRow sits on these seven

We hold ourselves to the same list, so here is where we fall short of it as well as meet it: one of the seven we fail, and one we only partly answer.

Our coach reads your last ten sessions — split times, watts, stroke rate, heart rate and the note you wrote on the row — and works out your current load, your fitness and fatigue trend, and how long since you last rowed. That is what it adapts to, and your conversation is stored, so the recent exchanges are in front of it next time.

On question 2 it fails the check: it knows how long since you last rowed and whether that leaves you in a short or a long recovery window, and its rules change the session when those say to — but a good answer asks why you missed the session before it re-plans, and this one does not. If you were ill or hurt, say so in the conversation. On question 3 the answer is only partial: what you logged is not what you told it, so tell it. On question 4 you can override in the conversation, and what you said stays there. On question 5 the boundary is training — it will talk about load, pacing and technique, and its system prompt carries a standing instruction to point you at a qualified coach or a physician, with a medical question going to the physician, because it is not one. On question 6 the record is your sessions and your chat. On question 7 it changes the session when the time since your last session and your recent load say to, and it has to show its working, because it is told to audit your load before it prescribes anything.

The one thing it will not do is pretend to be a coach who has watched you row. It has your monitor data. If you want a human who has seen your stroke, that is a different tool, and worth having. For how we handle your data, the Privacy Policy at /privacy is the only place to look.