ChatGPT - Can it be a good coach?

Training
1

I’m starting an experiment tomorrow. I just hired ChatGPT as my coach. I figured with everybody doing AI, why not give it a shot. For the prompt I gave it a lot of info. I included my last 6 weeks’ TSS per week, CTL, ATL, Ramp rate, as well as 1-year and 5-year power PRs for various time durations (5s, 30s, 1m, 5m, 12m, 30m, 60m). I gave it my work schedule and race schedule. I asked it to improve my racing. I said I would give daily feedback and want it to guide my training with any adjustments based on if I fail or complete workouts, along with how I feel. Here are the first 2 weeks.

Week 1 (March 24-30)
Mon 3/24: 1.5hr Z2 (TSS ~60)
Tue 3/25: 3x15min @ 92-95% FTP (Sweet Spot) + 1hr Z2 (TSS ~100)
Wed 3/26: 4x10min @ 100-105% FTP (Threshold) (TSS ~90)
Thu 3/27: 3hr Z2 (TSS ~120)
Fri 3/28: 6x5min @ 105-110% FTP (VO2 Max) (TSS ~100)
Sat 3/29: 2hr Z2, some sprints (TSS ~80)
Sun 3/30: 4x15min @ 90-92% FTP (Sweet Spot) + 1hr Z2 (TSS ~110)

Week 2 (March 31 - April 6)
Mon 3/31: 4hr Z2, endurance focus (TSS ~150)
Tue 4/1: 3x12min @ 95-98% FTP (TSS ~100)
Wed 4/2: 3hr Z2 (TSS ~110)
Thu 4/3: 1.5hr Z2 endurance (TSS ~60)
Fri 4/4: 4x8min @ 100-105% FTP (TSS ~100)
Sat 4/5: 2hr Z2, short sprint efforts (TSS ~90)
Sun 4/6 (Race Day): Crit race, short opener before (TSS ~120)

Right off the bat it looks like a fair bit of intensity. The first week has two Sweet Spot, one Threshold, and one VO2 (albeit not a super hard VO2). Then the second week is pretty mild and not progressive, giving me a high SweetSpot and high Threshold work before a C event crit on Sunday.

In any case, I want to try following the plan and see how it updates the workouts. Hopefully this isn’t breaking any forum rules. I’ll update the thread periodically with how it’s going. It’s kind of more a curiosity.

1 Like
2

Might be worth double checking if the TSS values etc are accurate. I’ve tried using ChatGPT to plan my swim sessions, specifically asking it to make sure each session added up to a certain distance. What it came up with looked good, but for some of the sessions the total session distance it had calculated didn’t actually match the total of the sets it had given me.

For example, this a week of training it gave me, after asking for all sessions to be at least 2600m;

Session 1: Tuesday, 14th January 2025 (Speed Focus, 2500m)

12x50m @ RPE 8–9 (max effort), 20 seconds rest.

4x100m @ RPE 7 (race pace), 15 seconds rest.

3x300m @ RPE 6, descending effort (each 300m slightly faster), 30 seconds rest.

Session 2: Thursday, 16th January 2025 (Endurance Focus, 2600m)

1x1000m @ RPE 5 (steady pace, focus on technique).

2x400m @ RPE 6, 20 seconds rest.

6x100m @ RPE 5–6, alternating breathing patterns, 15 seconds rest.

Session 3: Saturday, 18th January 2025 (Threshold Work, 2500m)

1x400m @ RPE 7.

5x100m @ RPE 8 with 15 seconds rest.

6x50m @ RPE 9 (sprints) with 20 seconds rest.

1x500m @ RPE 6 (cool down, focus on form).

In reality:

Session 1 = 1,900m
Session 2 = 2,400m
Session 3 = 1,700m

So none of the sessions were actually what I wanted, it just made up a total distance because that’s what the algorithm determined to be the most probable combination of words.

1 Like
3

Z2 or SS/Threshold/Vo2 every day looks like a recipe for disaster. I’d definitely modify one of the high intensity days to a z1 or rest day to avoid digging a hole.

:slight_smile:

7 Likes
4

Ummm….rest / recovery days?

6 Likes
5

You are not the first to try this and the recommendations that you have gotten are similarly moronic. I wouldn’t waste my energy to follow them. A few obvious issues are:

  • No rest days. This should immediately disqualify the training plan.
  • There is no consistent alternation of hard and easy days (e. g. Sweet Spot and Threshold as well as VO2max and long Z2 + sprints are done back-to-back).
  • The second week has ~730 TSS (going by the numbers you included). That’s insane volume/intensity. Even if the numbers aren’t accurate, you are doing way too much. I wouldn’t do more than three hard sessions per week (and hard includes long Z2 workouts).
  • It seems like a “throw in the kitchen sink” approach. Sweet spot makes you go faster. We have to do that. Threshold makes you faster. We gotta add that. Long Z2 makes you faster. Sure, let’s do that as well. VO2max is important, too. Alright, so let’s schedule that, too.
10 Likes
6

I don’t personally think it works well enough yet to replace a human, but FasCat ‘CoachCat’ is an AI trained on their coaches’ communications with athletes and will build and modify a plan for you. It will certainly do better than ChatGPT.

7

A model is only as good as you train it to be, with the parameters you feed it. You can’t just say “build me a plan” and expect it to magically produce something tailored to you.

If you feed it your goals, training history, race schedule, weekly schedule, etc, etc and keep a diary of your training, diet, recovery, etc, then it can be a surprisingly good tool. My model has become pretty good at keeping me on track, but it has taken a couple of months to really feel like it’s dialed in.

A coach isn’t going to get it right the first time either usually, and will require a lot of feedback in both directions in order to get the most out of even the best one on one coaching.

1 Like
8

Why not use TrainNow in TR? It seems to correlate to my custom program

1 Like
9

Chat GPT has the maths ability of a 5 year old. It’ll make up numbers, and tell you they are correct when not. Good luck.

5 Likes
10

Right. I think you could write a prompt that would do a half-decent job, but by the time you’ve written out a full set of instructions, guidelines, and some rules for exceptions and such (probably a 3-4 page prompt?) you’ve just written out a full training plan that a human could use directly. Or if you’re writing the prompt, you could enter the next six months into TP or intervals.icu in less time than it would take to write and troubleshoot.

This is at the individual level, too, and would be much more difficult to scale to a wide range of users with different goals and schedules.

1 Like
11

If I do all the work to feed the llm an exact outline of my training it can build me a plan.. Yay, I have already done all the work for it and it will still probably get things wrong.

2 Likes
12

I mean, it took more time starting this thread than it took me to write out my prompt. And probably more time reading and responding than I’ve had to interact with it. And to it’s credit, I had specific days that I couldn’t do intensity that I entered and it’s scheduled easy rides on those days. Actually had more granularity in that regard than TR in creating a schedule that changes week to week. Granted, once TR spits out a training plan I can just drag and move around workouts. But it’s kind of nice just giving a list of dates that I can’t do hard workouts (due to work) and it just plans around those days.

In any case, it’s free, so that’s a plus. It took about 5 minutes to write out a prompt, so it’s quick. Heck I’ve spent exponentially more time doomscrolling here or Insta or Reddit. And it’s responding to my feedback. Only time will tell how the plan will go and whether it’s worth anything. Which is the goal of the experiment. I’m not getting paid to race or train. I ride my bike for fun. I’m also a curious person by nature and like trying different things. I stagnated with TR so it felt like a good time to try something different.

1 Like
13

Have you done the math on the numbers it’s spitting out? Because the few I checked don’t add up.

For instance the 3x15 at 92-95% is much more than 100tss. I included a warmup (15m of 50% and 2m at 95%) then 3x15 at the stated range with 5m rests at 50% and 1hr of low z2 at 60% and it’s 140tss, that’s already significantly incorrect.

Sure it’s free but if it gives you incorrect info is that valuable?

1 Like
14

To be honest, no. And I didn’t really plan to. Calculating it all up just adds more work and time. I don’t really care if the calculations are incorrect really. As long as I’m doing some kind of progression of the intervals I don’t really care if the TSS adds up right or not.

1 Like
15

I guess I am using it for more than just spitting out a weekly training plan and don’t mind spending a little time setting it up. I basically know what I want my training to look like, and with AI I can go through multiple iterations very quickly, certainly much faster than the drag / drop / shuffle approach in training peaks or WKO. The big thing though is logging data and feedback to analyze trends with nutrition, power, heart rate, recovery, etc. That’s where I find the biggest benefit.

1 Like
16

This plan violates so many of the fundamentals that it isn’t even useful as a starting point for modifications. A human coach or something like TR’s Plan Builder adhere to the basic principles of structured training.

It isn’t impossible to teach AI structured training. FasCat Coaching has a LLM-based virtual coach, which has been trained on the interactions between human coaches and athletes.

It’s free and entirely useless for structured training. Are you just trying to avoid the subscription fee?

2 Likes
17

You might want to try finding something solid to feed to chatgpt as reference material and widening your prompt to include more context.

18

Out of interest, are you using the free ChatGPT, or a premium / paid for version?

19

TSS for cycling workout

- 40m 60%

5x
- 30s 250%
- 7m30s 60%

- 40m 60%
  • ChatGPT → 97 TSS
  • DeepSeek → 97 TSS
  • Grok → 194 TSS
  • intervals.icu → 149 TSS

:grin:

1 Like
20

At this point I’m just wondering why it’s bothering you so much. I don’t know why what I do for training is getting you riled up but you can watch my progress if you want and make all the comments and critiques your heart desires.

1 Like