The AI coach we (self-coached athletes) want

41

Really interesting thread. I’ve been lurking for a while and wanted to jump in because I’m building something that tries to solve exactly what you guys are talking about.

I’m working on a coaching app (iOS/Android/Web) that isn’t just another LLM chatbot spitting out generic advice. It’s a proper platform with structured workouts, zone calcs (Coggan, HR, CSS, etc.), and full load tracking (ATL/CTL/TSB). The AI coach works on approval—it proposes changes, tells you why, and you click yes or no. It doesn’t mess with your schedule without consent.

Under the hood, the coach has access to about 40 tools. It reads metrics, analyzes drift and power peaks, and has actual long-term memory. If you mention an injury or that you can never train on Tuesdays, it remembers that weeks later. I’ve already got integrations running with Strava, Intervals.icu (bi-directional), and Apple Health.

The point made earlier about plans failing when life gets messy really hit home. Generating a static plan is the easy part. The hard part is the daily grind—you slept 4 hours, you’re traveling, or you missed two sessions. That requires context, not just a recalculation.

The other issue I keep seeing is that AI coaches are too generic. They all pull from the same average internet knowledge. A TSB of -15 means “push” to one coach and “rest” to another. The AI shouldn’t be guessing; it should be following a specific philosophy.

That’s why I built a feature called Blueprints.

Think of it like a marketplace for coaching logic. Coaches or experienced athletes publish their methodology in a markdown format and you can select the blueprint for your training plan. It stops the AI from hallucinating generic advice and forces it to stick to that specific coach’s rules regarding periodization and recovery.

It also handles natural language notes. A Blueprint might say, “If the athlete misses a session, protect the Sunday long run at all costs.” So when your schedule blows up, the AI adapts exactly how that specific coach would, not how ChatGPT thinks it should.

The big plus of doing this as a marketplace is that you aren’t flying blind. You’ll be able to see ratings and popularity for each Blueprint, so you know which methods are actually working for other people before you commit to one.

A Maffetone Blueprint and a Daniels Blueprint will give you totally different advice on the same data, which is how it should be. It gives the AI a consistent voice.

Happy to show more if anyone is interested. Does the “Blueprint” concept sound like something you’d actually use, or am I overcomplicating it?

The results of my work are here: https://mytrainpal.app
Obviously still in early development days so please forgive if you find bugs, appreciate your feedback though.

42

Hi @Krzysztof , looks very interesting – but it is not what I want! :-o I do not want another platform, I want something like the logic you are describing in the intervals.icu platform😉

Now, don’t get me wrong, I will look at your platform when I find time (hard these days and months) and it might be quite good “by content”. However, generally I really look for avoiding every tool being a new platform, having a new frontend, etc. Coaching is a backend, and such backend coaches is what I am looking for with the – to me – familiar frontend dubbed i.icu;-)

43

I understand what you mean, in this case I feel like as tight integration as possible with intervals is key. The way I see it, if you want to stay in the intervals, but have some automatic training planning that you control, you can just enable auto sync with intervals in MyTrainPal. It will just serve then as an interface for AI to understand you, your lifestyle, your preferences and fitness to suggest workouts that will then land back on Intervals.

Actually if you don’t even want to touch any other app, you can set up scheduled prompts to execute every week, analyze last week’s data and update your plan in intervals.

1 Like
44

What I don’t want from an AI coach:

An average or generic training plan. It’s trivial to map out training sessions with progression, and adjust based on performance and how I feel, especially if it’s just a basic FTP build.

What could be useful:

Helping me understand the demands of a race and what I can do specifically top prepare for it.

Analyzing files after a race or event to figure out how I lost time and how I can improve. I spend a TON of time after every important race downloading the power files of everyone I can find, comparing segment times, trying to estimate weight, ftp, etc.

Consider individual response to training. Someone mentioned using athlete data to make predictions on performance. I think the problem with this at the population level is that individual response to training varies greatly, and I’m not sure if we really know much about whether an individual may response to one type of training vs another…like be a high responder to polarized training and low to threshold training. Every year I do a full season review and look at what kind of training I was doing leading up to my best (and worst) performances, but I just kind of eyeball it. Maybe an AI coach could do better.

Consider my preferences and the terrain available.

2 Likes
45

This sounds like an interesting idea. I will certainly explore it. But I have a question: you write of blueprints – which I take to encompass constraints, habits and also training philosophy. But how does the app handle goals and events? And if it does, can it handle multi-day events?
Thanks for your efforts!

46

You nailed it, that’s actually the other half of the equation I didn’t mention: Plans

You’re right that the Blueprint is just the philosophy and habits (the ‘how’). To actually apply that to a schedule, the app uses a Plan (the ‘what’ and ‘when’).

Basically, when you have a specific goal (like a Sub-3 Marathon) or a race date, that lives in the Plan. The AI then takes the rules from your Blueprint and maps them onto that specific timeline.

For events, I rely heavily on the integration with Intervals.icu. If you put a race there, it syncs over as a ‘fixed point.’

Regarding multi-day events. Yes, it can handle them. If you sync a 3-day stage race, the AI sees it as a sequence, not just isolated days. It recognizes the cumulative load and, based on the Blueprint’s logic, will usually force a deeper recovery block immediately after. It creates the context so the AI knows why you’re trashing yourself for 3 days straight and doesn’t try to stop you—but knows to rest you afterward.

Does that clarify things? I’d be curious to hear how you usually structure your multi-day blocks—it’s a complex problem and I’m always trying to refine the logic there.

47

It would be better to move further comments about your tool to a new thread in the “External Project” category. If every AI developer explains “their” tool in this thread and why it is better than others, that would obviously be the opposite of what the author of the original post wanted to read and discuss about.

7 Likes
48

“Analyzing files after a race or event to figure out how I lost time and how I can improve” → exactly this. I’ve start using extensive data just for a year now (since I joined Intervals) and discovered my BR is limiting me in races. That’s post training and race analysis. And training should be adapted to tackle these things. So far I haven’t found a great tool that is doing so.
(It bugs me a lot too that Zwift don’t record BR despite HRM is monitoring it, so I have 2 files that needs to combine when riding indoors )

1 Like
49

Understood, sorry for derailing the thread. My intention was, if we want something, I might as well just build it according to the observations people here made.

50

Yes, I (the “thread starter”) think there are two sides to this medal. It’s good that these discussion happen, also here, but then the thread should really not be diluted by technical details.
I’ll try to add a list of “spin-off threads” to the initial posting here – which I should update more anyway – so we connect all this a bit better.

But, it would be great – as @R2Tom points out – if the current thread is about philosophy and dreams on interacting with AI coaches, not on actual technical implementations nor on discussions and pathways that are orthogonal to the actual target atop😉

Thus, maybe also to future discussions, my 2 ct are that you should feel free to start off here for general discussions, but spawn into a separate “tool xyz” thread when it get’s technical or specific to the tool.

2 Likes
51

Hi @albauer what is the calculation behind Durability Index (DI)? I cant seem to find a unified definition online.
Thanks, Duppie