I did. I said, “using [blank] as a foundation for training… make me a plan that [blank] would give a client based on [their] training philosophy” and then continued with my prompt. I’ll let you decide on who I used.
???
I am surprised that this is what you took from my posts. You posted here to get feedback on the training plan Chat GPT has proposed, no?
Critiquing the proposed plan could be seen as a good exercise in learning training fundamentals such as periodization (which, importantly, includes rest), progressive overload, specificity and individualization. The proposed plan violates most of these basic principles on its face and its volume and intensity are unsustainable. (Edit: I’m not sure whose training philosophy you referenced in your prompt, but I doubt this is reflected in Chat GPT’s answer.) Other members have agreed with this assessment. 
I think if you want to use a Large Language Model to design your training plan, you have to use something more specialized like FasCat Coaching’s CoachCat. That’s been specifically trained on the years and years worth of interactions between their coaches and clients, so CoachCat — if it works well (I haven’t used it, so I don’t know either way, not meant as a dig) — will reflect an average of the training philosophy of FasCat Coaching.
A generic LLM like Chat GPT seems ill-suited and the proposed training plan reflects that.
5 Likes
Take away the second sweet spot and it’s identical to some TrainerRoad plans (1 VO2, 1 SweetSpot, 1 Threshold). And seeing as I’m giving it feedback day to day, one point that I’ve already given it is that the first week has a lot of intensity and the answer it gave is we’ll adjust the weekend intensity as needed based on how I feel. So already it is anticipating a possible change and swapping that second sweet spot to endurance, which would bring it to the same plan TR gives me. 
Seems like a reasonable place to make modifications.
I think the biggest part is going to be how it adapts and changes. Looking at a TR base plan, just look at how many questions are in this forum regarding the base plans. This is too easy, where’s all the endurance rides, where are my long rides. What TR does is adapt and change the plan. If you just did the stock plan as scheduled it probably wouldn’t do much for anybody. The gamechanger is AT and the analysis of completed/failed workouts and changing up the next workout. Only time will tell how well ChatGPT can do that. Which IS the experiment.
Not really, no. It was more just to document my experience with using ChatGPT as a coach. I figured I would get feedback but it wasn’t my main concern. Just sounded fun to try. And actually reading over the replies in the thread is entertaining and enlightening. I always like to hear different perspectives and different ways of thinking. I just thought it was funny that it seemed to get you worked up over the plan. Didn’t mean any offense.
1 Like
Still not a single rest / recovery day in 2 weeks…
2 Likes
That’s pretty much exactly what I was thinking in terms of a prompt. If I were to do one, I’d start with these sections:
- Persona/Goal - what’s the underpinning philosophy and what’s the end goal? Race crits? Long gravel races? Hill climbs? Raise FTP? TTE? 2-5 minute power? Lots of anaerobic work? Minimal/no anaerobic work? How do you get there? Polarized, pyramidal? Lots of VO2Max, SST/Threshold? Sprints?
- Periodization - how often are rest weeks? What’s the larger plan? Dedicated blocks with different focuses, like an SST/Threshold block or two, then a V02Max block? Lots of everything all the time?
- Volume - how many hours/week to start? what are you ramping up to? What’s the ramp rate?
- Weekly scheduling - what does a standard week look like? What are rest days? Is there enough recovery time over the course of the week?
- Exceptions - how does it handle missed workouts? How about a week or two off for illness or vacation? Does it replace workouts? Add a period to transition back before lots of intensity?
On the data analysis side, what kind of metrics are you looking for?
On a related note, it looks like there are integrations for Strava where you can connect it to an LLM, and I’ve been flirting with the idea of pulling them into a model to get a better sense of what my overall plan looks like during my best periods to see if it comes up with anything interesting. Not sure if the integration will still work considering Strava’s AI/ML integration policy, but is that only for third-party apps? I know volume very directly correlates with points my FTP is highest, but what other factors contribute? Time in various zones? Max intensity?
2 Likes
I have used ChatGPT to create plans as well. As always with these tools, getting the details right in the prompt makes a significant difference. In my case, the prompt always said to include 1 rest day per week and 1 recovery week every 4th week.
But I think this is an interesting experiment. Please keep us updated.
1 Like
I don’t generally schedule rest days. And it’s not uncommon to go 2-3 weeks before I take a day completely off the bike. Hasn’t been an issue in the past. Usually life will dictate a day off or I’ll just dial back an endurance ride and go short and/or easy that day. Which is what I plan to do. So I wasn’t too worried about it.
Edit: To clarify, I don’t think I’ve “scheduled” a rest day in 5 years. That’s not to say I don’t take rest days. Either my life (work, family, etc) will force me off the bike, or I just listen to my body and take a rest day when I need one (physically or mentally). I find much like intensity days sometimes, rest days don’t fall on days that make sense. There are days that I get antsy if I don’t ride, so mentally they just mess with my head. Not to mention that the day after a rest day I usually feel like sh!t so I’ve found that an easy 30 minute spin puts me in a way better place to hit workouts. I still take full recovery days off the bike, but they’re pretty rare. Just my experience over the years so it might not work for everybody.
1 Like
Mine started with having it analyze a coaching philosophy. I said “learn everything about [coach] and their overall training philosophy.
Then I gave it my stats: CTL, ATL, ramp rate, last 8 weeks’ TSS and hours. I also fed it my PRs for a handful of durations. I entered my goals for the year and beyond. And then fed it a basic daily schedule of work and life, as well as specific days that I couldn’t do intensity (day after a night shift). I entered all of the races I had planned for the year so far (date and type of race), and put in my focus races.
So I said, get me stronger and faster with those constraints and background, using [coach] as the basis.
It gave me a few months of day by day training, with general outlines of the future past that with the idea that it would change.
I actually then gave it the prompt, “learn [different coach] and have that coach critique the plan.” The main critique was too much intensity. So then it said it will dial it back as needed.
2 Likes
I agree. The prompt is the key. Can have huge differences. I try to be as specific as possible. I feel like you need to give it fences to keep it on track.
Nice, that’s a pretty cool exercise. I’m super curious how the prompt evolves over time.
1 Like
My general rule of thumb with LLMs is to not use them for anything where I need factual correctness if I’m not in a position to validate that correctness or if doing so would take me longer than doing the task myself, and this does seem like a smell that the response it’s giving is answer-shaped but not correct. With all the cycling training material it probably has in its training data it’s surprising to see it not suggest even a single rest day, so I’d question the rest of the output as well, but maybe that’s just me.
6 Likes
I don’t think this can be overemphasized. LLMs have an amazing ability to be confidently, stridently, absolutely wrong. I work with AI in healthcare data, and there’s no such thing as too many variations of “Stop making shit up!” in a prompt. “Do not make anything up. Do not condense. Do not summarize. Refer only to the included data, do not make any inferences or extrapolations. For the love of god and all that is holy, stop making shit up.”
12 Likes
Yup.
Overall, the basics of the plan will be “fine”. A little intensity, a little endurance, a little in-between. Mix and match as you’d like.
Will it “work”? If you define “work” as improving fitness, then the answer is probably “yes”. Will it be “optimal”? You’ll never know. My initial impression is that it is too heavy on intensity and doesn’t provide enough recovery, which are some pretty basic mistakes. As long as you have the experience to adjust for that, no issues. But if someone tries that plan without a decent knowledge of training principles, they are likely heading towards a cliff.
3 Likes
I sort of answered this above. But that is kind of from the prompt. I didn’t ask it to give me any rest days. Another poster said they specifically asked for rest days every so often. It gives you what you put in. I explained my reasoning above regarding not scheduling specific rest days so it’s not that it forgot them, I didn’t ask for them.
What I’m saying is that because rest days and recovery weeks are common in training plans and training literature in general, it should have given them to you, so the fact that it didn’t add any makes me question what data it’s trained on, and by extension the correctness of the output.
2 Likes
I just went back through TrainingPeaks and I’ve had 3 days off the bike so far in 2025. One was flying to an event so I was traveling. Another was the day after a 3-day omnium and I needed a day off. The third was when I had a particularly long day at work and just felt like plopping on the couch rather than ride. Apart from those, I don’t remember any days that I felt like I needed a rest day. Like I said, I might get on the bike and take it easier than planned or just do a quick spin. But I rarely take full days off the bike.
1 Like
It has recovery weeks later on with the typical drop in TSS and intensity. But it doesn’t have rest days because I didn’t want them. Week 6 of the plan is a recovery week. So longer than maybe most plans if you’re used to 3-1 or 4-1. I only posted the first 2 weeks because it’s a wall of text to post 6 weeks out. And because I figure it will change a lot by then so I didn’t want to post it all if it changes.
1 Like
AI can be like a trainable monkey at times. You can train it to intentionally give wrong answers, just as you can train it to give you correct answers. See below.
Until your post, I assumed that everyone knew that you HAVE to further train these pre-trained models in order for them to be very useful. What fascat is doing is entry level stuff. If you really want to see something wild (and this is just scratching the surface), you can build a full stack app that will link to strava, download all of your workouts, and apply whatever processing and analysis you want… with AI, for free. (Not chatgpt) Someone that knows what they are doing could have this functional within a day. This is still the ground floor for what’s to come…
1 Like
That is pretty good, I usually prompt it saying something like
"Act as a top doctor, sports scientist and cycling coach, using the following documents weighed and assessed for their scientific rigor as per your judgement …[pdfs] as well as the common public record to answer the following questions, do not be afraid to propose new reference material published in the last 10 years or 20 years in the case that rigor or foundational basis can be established ::
context :: I am a x year old cyclist with y weight and z training history, I have k race results, a compound score of j and I want to improve n part of my performance. My biggest haters would say I suck at M and on my best days I ride like xy
With the right mix of papers fed in usually I use at least
Assessment of Metabolic Flexibility by Means of Measuring Blood.pdf (1.3 MB)
criticalpowerandwprime.pdf (670.5 KB)
maintaining_power_output_with_accumulating_levels.12.pdf (354.4 KB)
workload characteristics of elite cycling.pdf (307.6 KB)
Just have to stay on your toes and make sure to ask the AI to play devils advocate if it seems to be agreeing with you far too often. Chatgpt has a habit of forming a bit of an echo chamber where if you speak plainly to it after prompting it will start to interpret your talk with too much weight, so for example if it starts saying you should work on X and you reply with “oh I do good at X” then it will take that as a factual statement and trust you which is kind of defeating the point because as you may know part of coaching is having someone defeat your own freudian superego and tell you that you actually suck at many things you think you are good at. 
Sometimes I like to also find older books about stoicism or other philosophy by logic texts and ask the AI to find overlaps between cycling race dynamics and the fall of rome or whatever, synthesis is a strength of the LLM so it does a pretty amazing job at that kind of thing and can make some very thought provoking parallels. But as always it’s a fine line between making real progress by means of synthesis vs the AI turning into an echo chamber so you really have to watch the way in which you prompt it.
1 Like
If it is getting such a simple thing wrong as calculating tss why would one trust it to get everything else correct?
1 Like