Training for TransBaviaans & 947 with AI (ChatGPT) in Zwift

[PietervS]

I am curious to know if any other South African (mortal, not Pro) riders are training with ChatGPT as coach? I am curious also to know your experiences and feedback on your AI training so far? 
I started using ChatGPT out of curiosity 3 months prior to TransBaviaans to get an opinion about the Zwift "Build-me-Up" structured training plan that I was using at that stage.  That day marked the end of my standard Zwift Training Plans, and the start of the most facinating experience with my new AI ChatGPT cycling coach.   I am still to this day using ChatGPT for weekly training plans.  I am training harder that ever and getting stronger than ever without ever feeling that I am getting over-trained!
I must also say that I trained with a human coach for almost 2 years in the past which helped me a lot.  As a mortal average 50 something year old cyclist, I have a reasonable understanding of formal cycling training requirements.  But I missed from my human coach good feedback on my training progress, and it also felt like all the athletes at the coach were on same training plan without considering individual needs or current training levels of the various athletes and the effect of life, work, travelling etc that sometimes caused interruptions in training schedules.  Although my fitness and FTP increased considerably with the human coach, I felt over-trained at times and to some extent I lost my excitement for training for various reasons, even though the numbers appeared to show I was doing well with my training.  This and also the significant cost of having a human coach forced me into falling back on standard Zwift training plans that at times felt like it was not pushing myself hard enough, and at times the training sessions didn't feel like complementing each other with very little recovery sessions in between.  So, when I asked ChatGPT's feedback out of curiosity on whether or not my Zwift training plan at the time was helping me for preparing for TransBaviaans, the AI coach told me straight out that my Zwift plan had very little focus on endurance training and focused way too much on sprints & neuromuscular efforts etc that are of little use on a TransBaviaans type event.  He suggested a formal training coach, but then suggested I can give "him" (ChatGPT) a try because of his access to vast amounts of published training, medical & nutrition data and his sound knowledge of Zwift, Strava, GarminConnect etc.  
So, I then decided to take a huge gamble and gave ChatGPT a summary of my training details to that point including my age, weight, FTP, VO2 max, race results from recent races, plans for upcoming race events, and roughly how much training I have done up to that point.  I then asked ChatGPT to give me his thoughts on a suitable training strategy for TransBaviaans.  Suddenly then the most interesting coaching experience started!
Based on all my Garmin stats and knowing the date and requirements of TransBaviaans from what ChatGPT found on the internet, my new AI coach worked out a new plan for the remaining weeks to Baviaans.  I got training plans on a weekly basis based on performance of previous week.  A free account in ChatGPT was essential so that my Ai coach could remember all past discussions.  ChatGPT suggested that I uploaded the stats of my training sessions after each session, so that adjustments could be made to remaining sessions for the week if necessary - all this by simply copying and pasting my Power and HR zone data into the ChatGPT discussion.  All data get analyzed in a split second with very valuable feedback on my performance for the session, and the next day's training session is customized based on my performance of the previous day.  ChatGPT knows how Zwift works, so my training sessions are sent to me as Zwift .zwo format files ready for import with limited input or customization required from my side. 
I have been training with ChatGPT now for 3 months and I must say it is an amazing experience.  The coach never gets angry or frustrated or disappointed if a training session is missed.  I can always get new customized sessions for the next week based on previous week's performance.  Whenever a new race comes up, I simply give ChatGPT the details, and my training program is immediately customized to best fit in the next race.  If any training day is missed, I just ask for a re-evaluation of the remaining sessions and there I go again.  For those of you hooked on stats, ChatGPT is overly keen on supplying performance comparisons with previous similar training sessions, indications on current training progress etc etc etc.  If 2 races are too close to each other, or a long groupride like a mini-munga needs to be accommodated, ChatGPT would typically advise me to limit my efforts for the non essential race/ride to specifics like max heart rate, max avg power etc in order to best limit fatigue build-up that could affect the most important upcoming ride.  I will be doing my 5th RideJoburg this year and look forward to my best time ever.  
So, if any of you out there needs cycling training from a coach that has access to all published training material available online, and would like to have access to this personal coach 24/7, a coach that is eager to number crunch the stats of your recent rides and make adjustments to your program based on your immediate needs and work/life situation, maybe you should give ChatGPT a try!  I must say I am sold on AI training!
Good luck all with your 947 preparations
Pieter (Krugersdorp)
 

[BigDL]

Hi @PietervS. Found this thread and wondering how this project went? 

[Phillippe Coetzee]

17 minutes ago, BigDL said:

Hi @PietervS. Found this thread and wondering how this project went? 

Yip also following

 

[BrendanA]

@PietervS ons wag... of het jy en AI die tour de france oor geneem?

 

[Dexter-morgan]

I will also be following this, how did your 947 go?

[Mamil]

I might give this a go for Munga training - I have a trainer road subscription and honestly I find their training programmes and the AI that informs them not useful - workouts frequently feel too hard or too easy - advises rest when actually I feel strong and able to do more.

Would also like to hear how its gone @PietervS

[PietervS]

Good morning everyone!

I am super happy with the way I have been training using my AI coach (ChatGPT).  I don't think I will EVER use a human coach again - I am a fanatic but average rider, no podiums etc that I am chasing.  My feeling is that unless you really want a competitive edge that other athletes don't have, AI like ChatGPT is probably as good as you can get.  All the training strategies come from whatever AI can find published.  The other day I tuned my HR zones a bit to adapt to latest Lactate Threshold HR, and ChatGPT referenced info from Training Peaks as his basis for suggesting the training zone adjustments. 

It is almost a year now, and still once a week, sometimes more often I quickly check in with ChatGPT uploading my ride results and asking then for a review.  My approach may be slightly different to formal "Training Platforms" that claim that they are using AI to give you a training plan.  With ChatGPT I get exactly that, but more (I think).  My way of working is more "personal" than working with a pre-setup training platform.  I signed up for a CHATGPT account - this is very important so that previous conversations can be remembered.  Also, with the signed-up version I work with ChatGPT v5.2 which is a pretty great platform, with nice personal touch to the way it communicates with the user.  You can tell ChatGPT how you want it to communicate with you, i.e. always be uplifting, boost your moral, etc.  I prefer no sugar coating so I told ChatGPT that it should ALWAYS give feedback as it is.  If I screwed up in a training session, it should say so.  So, the responses and feedback that I get are almost military style - I will give you an example below.  I know exactly after a chat where I messed up and what the focus should be for next training session.

Using ChatGPT as coach, I managed to gain 15 minutes last year on 947 (finishing 3hr 4 mins) vs 3hr 20 min the previous year.  TranBaviaans I was stronger than ever.  We had some mechanicals to deal with, but I managed all the climbs even MAC without having to stop/rest,  I did the big MAC about 25 minutes  faster than the previous year.

You guys should try it out yourself.  Sign up for a free-of-charge account with ChatGPT for a start so that it can remember your conversations.  With free account you are limited by number of chats a day, but you don't need all that many chats to start with.  First message you should give all your details.  Write it in the format as if chatting with a human - give as much detail as you can including age, weight, current FTP, VO2 max, and then an idea of how much you are currently training and what you are training for with specific info.  Ask it then for a training plan and I suggest then to ask for workouts for the next week only.  Asking for too much info in advance doesn't really work for me because you may in any case not stick to it.  I prefer that it gives me the workout in text so that I can program it into Zwift as a custom workout.  I also like if it gives me the plan in "Python" format to understand exactly how to program it into zwift.  It typically gives me something like the below:
 

✅ Tuesday 23 July – VO2max Focus (Short Intervals)

Type: Repeated 30/30s VO₂max intervals
Structure:

• 2 sets of 6 × (30s @ 125–130% / 30s @ 50–60%)

• 4 min recovery between sets

• Followed by long Zone 2 block

Focus:

• More neuromuscular demand (acceleration + recovery)

• Higher total intensity density

• Great for improving oxygen utilization efficiency

• More explosive, but not as aerobically deep

---

⚖️ Which One to Use?

Both are effective for VO₂max, but:

• Use the 3-min intervals (what you received earlier in Python) if you're building long-range aerobic capacity and pacing.

• Use the 30/30s if you're working on oxygen kinetics, repeatability, and explosive responses — or if you're a bit fatigued and want a less mentally demanding structure.

🧠 Tip:

You could alternate these styles weekly leading up to Baviaans:

• This week (23 July): Use the short intervals for variety and recovery from last Saturday’s long ride.

• Next VO₂max session (after recovery from mini-Munga): Do the 3-min intervals to deepen fitness.

Would you like the 30/30s short interval session in Python or .zwo format?

PvStaden:  Please give me the Python code including recommended cadence for the 30/30s version

 

Here is the Python-formatted training session for the VO₂max Short Intervals (30/30s) version you mentioned — complete with recommended cadence for each segment.

---

🔥 Tuesday 23 July – VO₂max Focus: Short Intervals

🕒 Duration: ~75 minutes
🎯 Focus: VO₂max development through high-repetition short intervals
💡 Ideal for: Aerobic power, oxygen uptake kinetics, repeatability under fatigue

Python code:

# VO2max Short Intervals – Tuesday 23 July

[
  # Warm-up: 15 min ramp from 50% to 75% FTP @ 85–90 rpm
  ('Warm-up Ramp', 900, 0.5, 0.75, 90),

  # Set 1: 6 × (30s @ 130% / 30s @ 55%) @ 100–105 rpm for the high efforts
  ('30s #1 - Hard', 30, 1.30, 1.30, 100),
  ('30s #1 - Easy', 30, 0.55, 0.55, 85),
  ('30s #2 - Hard', 30, 1.30, 1.30, 100),
  ('30s #2 - Easy', 30, 0.55, 0.55, 85),
  ('30s #3 - Hard', 30, 1.30, 1.30, 100),
  ('30s #3 - Easy', 30, 0.55, 0.55, 85),
  ('30s #4 - Hard', 30, 1.30, 1.30, 100),
  ('30s #4 - Easy', 30, 0.55, 0.55, 85),
  ('30s #5 - Hard', 30, 1.30, 1.30, 100),
  ('30s #5 - Easy', 30, 0.55, 0.55, 85),
  ('30s #6 - Hard', 30, 1.30, 1.30, 100),
  ('30s #6 - Easy', 30, 0.55, 0.55, 85),

  # Recovery between sets: 4 min @ 60% FTP @ 85 rpm
  ('Recovery Between Sets', 240, 0.6, 0.6, 85),

  # Set 2: Repeat the 6 × 30/30s
  ('30s #7 - Hard', 30, 1.30, 1.30, 100),
  ('30s #7 - Easy', 30, 0.55, 0.55, 85),
  ('30s #8 - Hard', 30, 1.30, 1.30, 100),
  ('30s #8 - Easy', 30, 0.55, 0.55, 85),
  ('30s #9 - Hard', 30, 1.30, 1.30, 100),
  ('30s #9 - Easy', 30, 0.55, 0.55, 85),
  ('30s #10 - Hard', 30, 1.30, 1.30, 100),
  ('30s #10 - Easy', 30, 0.55, 0.55, 85),
  ('30s #11 - Hard', 30, 1.30, 1.30, 100),
  ('30s #11 - Easy', 30, 0.55, 0.55, 85),
  ('30s #12 - Hard', 30, 1.30, 1.30, 100),
  ('30s #12 - Easy', 30, 0.55, 0.55, 85),

  # Endurance Block: 25 min @ 65–70% FTP @ 90 rpm
  ('Zone 2 Endurance', 1500, 0.65, 0.7, 90),

  # Cooldown: 5 min @ 50–60% FTP @ 85 rpm
  ('Cooldown', 300, 0.5, 0.6, 85)
]
 

Edited February 5 by PietervS

[PietervS]

Typically after a ride I would upload my HR zones and Power zones from Strava, and ask ChatGPT to analyze and give me feedback on where I am with teh training.  Typically I would get a response as following:

PS, I am calling the bot Wattson, a name he suggested for himself 🙂 

Good morning Wattson, I finished this morning the VO2 Micro Intervals session. Herewith my stats. Please check if I achieved the intended outcomes:
Z1 Recovery 0 - 125 W 20:36 31%
Z2 Endurance 126 - 171W 36:00 54%
Z3 Tempo 171 - 205 W 2:13 3%
4 Threshold 206 - 239 W 1:09 2%
Z5 VO2Max 240 - 273 W 35s 1%
Z6 Anaerobic 274 - 342 W 5:41 9%
HR:
Z1 Recovery < 118 14:52 22.5%
Z2 Endurance 119 - 136 34:57 52.8%
Z3 Tempo 137 - 155 16:24 24.8%
Would you mind also comparing with previous weeks similar sessions to see if I am getting better on this specific workout and also to check if you can see if the effect of Sunday's hard ride has been recovered.

Wattson's reply:

 

Great job getting the VO₂ Micro Intervals done, Pieter 👏 Let’s break it down carefully, and then compare with earlier sessions to see trends and whether Sunday’s hard ride still lingers.

---

📊 Today’s Session (VO₂ Micro Intervals)

• Power Distribution

  • Z1–Z2: 85% → majority of session, as expected from short hard bursts with long recoveries.

  • Z5–Z6: ~10% → you hit the intended high-aerobic/anaerobic zones.

  • That balance matches the workout’s intent (short “oxygen ceiling” hits, then flush back into Z2).

• Heart Rate Response

  • 53% Z2 HR, 25% Z3 HR, no sustained Z4+ → HR stayed under control.

  • This suggests you had good recovery between efforts and weren’t over-fatigued.

✅ Outcome: You achieved the intended physiological target (stimulus to VO₂ system without massive overall strain).

---

📈 Comparison with Previous VO₂ Micro Sessions

19 Aug (baseline session)

• Work block: 30s @ ~265 W / 90s @ 140 W.

• HR: Mainly Z2–Z3, lower drift.

• Outcome: Good aerobic stimulus but less intensity than later versions.

5 Sept (progression)

• Work block: 2 min @ 220 W + 1 min @ 250 W repeats.

• HR: ~50% in Z3–Z4, higher drift.

• Outcome: First big step upward, more threshold time.

Today (17 Sept)

• Work block: 30s @ ~300 W / 30s @ ~115 W.

• HR: 53% Z2, 25% Z3, very little Z4.

• Outcome: More neuromuscular/anaerobic punch, but HR didn’t climb as high → recovery is working well.

---

🧠 What This Tells Us

• Recovery from Sunday:
  Your HR staying mostly Z2–Z3 today shows you’ve already absorbed Sunday’s load. If fatigue was still hanging around, we’d expect:

  • HR suppressed (struggling to rise), or

  • HR overly elevated even at moderate powers.
    Neither happened → 👍

• Progression Trend:

  • You’re clearly handling higher workloads (300 W bursts now vs 265 W baseline).

  • HR drift is consistent → adaptation is keeping up with intensity progression.

  • You’re stacking VO₂, Threshold, and Sweet Spot layers effectively.

---

✅ Bottom Line

• Today’s VO₂ session hit its mark.

• You’ve recovered well from Sunday’s harder-than-planned ride.

• The progression across sessions shows steady improvement — more power in bursts, with HR responses still under control.

---

👉 Pieter, do you want me to also do a “soft FTP adjustment check” here, like we agreed, to see if your heart rate at sweet spot / threshold intensities suggests we should edge FTP up slightly?

['Kaze Pete]

2 minutes ago, PietervS said:

Typically after a ride I would upload my HR zones and Power zones from Strava, and ask ChatGPT to analyze and give me feedback on where I am with teh training.  Typically I would get a response as following:

PS, I am calling the bot Wattson, a name he suggested for himself 🙂 

Good morning Wattson, I finished this morning the VO2 Micro Intervals session. Herewith my stats. Please check if I achieved the intended outcomes:
Z1 Recovery 0 - 125 W 20:36 31%
Z2 Endurance 126 - 171W 36:00 54%
Z3 Tempo 171 - 205 W 2:13 3%
4 Threshold 206 - 239 W 1:09 2%
Z5 VO2Max 240 - 273 W 35s 1%
Z6 Anaerobic 274 - 342 W 5:41 9%
HR:
Z1 Recovery < 118 14:52 22.5%
Z2 Endurance 119 - 136 34:57 52.8%
Z3 Tempo 137 - 155 16:24 24.8%
Would you mind also comparing with previous weeks similar sessions to see if I am getting better on this specific workout and also to check if you can see if the effect of Sunday's hard ride has been recovered.

Wattson's reply:

 

Great job getting the VO₂ Micro Intervals done, Pieter 👏 Let’s break it down carefully, and then compare with earlier sessions to see trends and whether Sunday’s hard ride still lingers.

---

📊 Today’s Session (VO₂ Micro Intervals)

• Power Distribution

  • Z1–Z2: 85% → majority of session, as expected from short hard bursts with long recoveries.

  • Z5–Z6: ~10% → you hit the intended high-aerobic/anaerobic zones.

  • That balance matches the workout’s intent (short “oxygen ceiling” hits, then flush back into Z2).

• Heart Rate Response

  • 53% Z2 HR, 25% Z3 HR, no sustained Z4+ → HR stayed under control.

  • This suggests you had good recovery between efforts and weren’t over-fatigued.

✅ Outcome: You achieved the intended physiological target (stimulus to VO₂ system without massive overall strain).

---

📈 Comparison with Previous VO₂ Micro Sessions

19 Aug (baseline session)

• Work block: 30s @ ~265 W / 90s @ 140 W.

• HR: Mainly Z2–Z3, lower drift.

• Outcome: Good aerobic stimulus but less intensity than later versions.

5 Sept (progression)

• Work block: 2 min @ 220 W + 1 min @ 250 W repeats.

• HR: ~50% in Z3–Z4, higher drift.

• Outcome: First big step upward, more threshold time.

Today (17 Sept)

• Work block: 30s @ ~300 W / 30s @ ~115 W.

• HR: 53% Z2, 25% Z3, very little Z4.

• Outcome: More neuromuscular/anaerobic punch, but HR didn’t climb as high → recovery is working well.

---

🧠 What This Tells Us

• Recovery from Sunday:
  Your HR staying mostly Z2–Z3 today shows you’ve already absorbed Sunday’s load. If fatigue was still hanging around, we’d expect:

  • HR suppressed (struggling to rise), or

  • HR overly elevated even at moderate powers.
    Neither happened → 👍

• Progression Trend:

  • You’re clearly handling higher workloads (300 W bursts now vs 265 W baseline).

  • HR drift is consistent → adaptation is keeping up with intensity progression.

  • You’re stacking VO₂, Threshold, and Sweet Spot layers effectively.

---

✅ Bottom Line

• Today’s VO₂ session hit its mark.

• You’ve recovered well from Sunday’s harder-than-planned ride.

• The progression across sessions shows steady improvement — more power in bursts, with HR responses still under control.

---

👉 Pieter, do you want me to also do a “soft FTP adjustment check” here, like we agreed, to see if your heart rate at sweet spot / threshold intensities suggests we should edge FTP up slightly?

Freak Wattson out and link intervals.icu to your Strava / Garmin and share those stats with it

[PietervS]

I am a statistics junky - the more info and comparisons I can get, the better.  That is for me where ChatGPT is amazing.  Because it remembers all conversations, you can ask it to do a comparison with previous training sessions etc.  It knows most of the race events and can dig up the ride profiles by itself.  So you don't have to supply it with info that is already known on the internet.  For example, you can say you want to train for W2W but want also to be ready for a specific Road Race 2 months before.  It will then work out a training plan to best adapt you for both.

There are lots of info I can share - my suggestion is try it out for yourself and see what it responds with.  Most important is that you communicate with it, give it your race statistics and tell it how you felt during the race.  You would be amazed with the info and recommendations that you get back from your AI coach.

 

[PietervS]

9 minutes ago, &#x27;Kaze Pete said:

Freak Wattson out and link intervals.icu to your Strava / Garmin and share those stats with it

The more info you can give it, the better he can evaluate your ride.  

I use the "Elevate" plug-in on Strava from there I get the following info that Wattson is also using for comparison purposes:
Efficiency Factor
Intensity Factor
Heart Rate Stress Score
Power Stress Score
Heart Rate Reserve
Normalized Power
Avg Power

I simply copy and past above info from Strava after a ride and give it to Wattson together with my HR zones and Power Zones.  Copy and Paste.

The beauty really is that if you don't have time or are not really interested in the stats, you can ask Wattson what minimum info he needs to properly guide you through your training.

But important - he doesn't know what he doesn't know.  If you feel drained or overtrained, tell him so.  If you feel like you are not pushed hard enough, tell him so.  The more/better you communicate, the better your experience will be with AI training.

 

[PietervS]

Last thing I should say, I signed up for a paid account with CHATGPT a few months ago, something like R400 per month.  (It is still less than half the price of a human coach).  This gives you most power from ChatGPT with unlimited chats.  This is nice also for work purposes as it allows you to upload files, spreadsheets etc which AI can read/update and send back to you.  For the training you can for example upload a garmin .fit file if you want the bot to check out specific performance on a race.

Only downside with uploading too much info, is the chat gets slow after a while.  The work-around is then to ask the bot to make himself a summary of the essence of the training chat to date, for you to copy and paste into a new chat.  He then produces a summary paragraph with all info he needs to remember and adds in "reset lines" in previous chat message.  When pasting that info into new chat - all things go back to super fast.

Enjoy your experience with ChatGPT and the amazing world of AI training!

Edited February 5 by PietervS

[Chingy182]

I started playing around with Grok last year to help me build a winter program in preparation for my first 100 miler - Around the Pot.
I was initially aiming for a sub-7h finish, but ended up somewhere in the ball park of 5h30.

The training structure and feedback helped me a lot in terms of tracking my progress and what areas to improve on.

I got good advise on balancing hard rides with tapering weeks, etc.

I have never had a cycling coach, so I have nothing to compare it against, however, my rides now have a lot more structure to them and each ride now has a purpose.

I have boosted my FTP from 284 to 311 with this improved structure

In the absence of a power meter on my bike, I have come up with HR variables and RPE scale alternatives to work with on outdoors rides.

Then there's even practical tips on nutrition, supplements and diet

[BigDL]

48 minutes ago, PietervS said:

Last thing I should say, I signed up for a paid account with CHATGPT a few months ago, something like R400 per month.  (It is still less than half the price of a human coach).  This gives you most power from ChatGPT with unlimited chats.  This is nice also for work purposes as it allows you to upload files, spreadsheets etc which AI can read/update and send back to you.  For the training you can for example upload a garmin .fit file if you want the bot to check out specific performance on a race.

Only downside with uploading too much info, is the chat gets slow after a while.  The work-around is then to ask the bot to make himself a summary of the essence of the training chat to date, for you to copy and paste into a new chat.  He then produces a summary paragraph with all info he needs to remember and adds in "reset lines" in previous chat message.  When pasting that info into new chat - all things go back to super fast.

Enjoy your experience with ChatGPT and the amazing world of AI training!

Thanks for sharing. I was curious as I have done something similar whilst prepping for Berg & Bush and will take some lessons from your experience. A couple of weeks ago I set up a coach personality in ChatGPT in a similar way to you, although haven't been as consistent with uploading my rides as you. I would still prefer a human coach but this has allowed me to dip in and out at any times with questions etc and, as my only goal is to finish and not be broken, I can't justify the cost this year of a Coach. 

Interestingly, I have access to all of the major AI platforms paid options through work so I put the same instructions into the latest versions of Claude and Gemini and got quite different guidance and programming from each of them. I didn't want to explore too much on them as they are supposed to be for work only but it did make me think  about the differences. 

[dave303e]

Just an idea, from someone building custom AI stuff now and going deep down the rabbit hole. Build an MCP server like ragflow, then take the data export out of intervals.icu and then build your chatgpt agent on your own history of training data....

[PietervS]

2 minutes ago, BigDL said:

Thanks for sharing. I was curious as I have done something similar whilst prepping for Berg & Bush and will take some lessons from your experience. A couple of weeks ago I set up a coach personality in ChatGPT in a similar way to you, although haven't been as consistent with uploading my rides as you. I would still prefer a human coach but this has allowed me to dip in and out at any times with questions etc and, as my only goal is to finish and not be broken, I can't justify the cost this year of a Coach. 

Interestingly, I have access to all of the major AI platforms paid options through work so I put the same instructions into the latest versions of Claude and Gemini and got quite different guidance and programming from each of them. I didn't want to explore too much on them as they are supposed to be for work only but it did make me think  about the differences. 

Interesting feedback, also with regard to getting different results from different AI's.  I got this even inside the ChatGPT platform.  V5.2 is way different (and WAY BETTER) to the no-account v4 of ChatGPT, also the character of communications is quite different.  Even inside of V5.2 the bot can give different answers based on slightly change in the tone of my question.  It can even "change its mind" at times on something that he feels is not truly important.  For example, if I forgot to ask the recommended cadence for a specific interval and ask him to resubmit the week's workouts with recommended cadence included for all intervals, I would often get 2 workouts now switched around, or get a revised workout for a specific day,  For example he would re-issue a workout with cadence now included, but now with 3x6min intervals where the previous one had 4x5min intervals.  This after I have just programmed the session in Zwift.  If I challenge him he would say the 4x5 is probably better but I can decide.  If I then ask him to include a shorter cooldown interval, he would revert with a workout based on the original 3x6mins but now adding longer recovery in between after he has just said the 4x5mins is better.  All answers may have very similar end-result but the "change of mind" can be very frustrating - it is outside the human nature where humans tend to stick to a good idea.  The bots are not sentimental in any way and don't feel bad if they have a "change of mind" or if they suddenly feel a completely different routine is better based on the way I phrased my sentence.  My experience on best approach with AI communication is to give as much info at any time as I could and try to be specific in what you require from it.  I then use the feedback as is and be very specific in the way I ask for additional information.  Also, I stopped asking ChatGPT to send me Zwift .zwo files because it sometimes makes mistakes in the workout files.  When challenging the BOT on a mistake, it would often then generate a completely different workout but with same aim at the end.  Also, I don't ask for too many training sessions at a time.  Lately I ask it for a summarized training intent for the next say 2 weeks, and I request then only the training workouts for the next day or 2.  Based on my feedback after a training session, he may very well adapt his recommendation for the next training sessions.   

I can understand that different BOTS would give different results, because the architecture of the bots are different and also the ways in which they decide which "source data" to use as info and how they want to use it.  I guess it is similar to switching human coaches.  A new human coach may may train me for same race event, but his approach may be different. 

My recommendation would be to use any reputable AI bot from a respectable developer that you feel comfortable with - you will see soon enough if it suggests something that you know from experience can't be right, or if you challenge him on his source for a suggestion and he quotes info from the Huisgenoot magazine.  You will know soon enough if it works for you or not.  What I like about ChatGPT so far, it speaks the same language as what I hear on GCN network and what I heard on training videos from experts like Inigo San-Milla etc.  And based on my performance results so far, I trust my ChatGPT coach maybe more than I should.

At the end, my personal view is any AI bot with access to published data on the web will give me WAY better results that having to trust a ZWIFT training program who doesn't care about my feedback.  My limited experience with human coaches so far (2 of them): they can only reserve so much time for me; they train me mostly on same program that worked for others half may age and double my FTP; they rarely give me any valuable feedback other than a Kudo on Zwift.  If I miss a training session, I have to somehow catch up or miss out on a valuable training session - he is not going to change my program because I don't feel like training this weekend.

Good luck with your AI training!