That is what I’m saying. And it adjusts based on that. Just like talking to a real person. The adjustments I mentioned are what it has given me based on what my day looks like.
1 Like
As long as you’re not saying the above bit, it’s all good
Did you try with the deep research functionality (became available for me today)?
This is something that you can do in the same chat with ChatGPT and one of the tricks to getting better results.
You ask it to generate x (code function, training program, homework) then the next chat prompt should be, review what you’ve just given me and tell me some improvements.
For instance I prompted mistral-nemo-instruct (a locally running LLM)
As a senior coach, doctor and cyclist give a 40yr old male with extensive experience of cycling and training a 4 week FTP builder plan. Based on 10h a week availability, 1 rest day a week and one long ride at the weekends. Current FTP is 300W.
Weekly Schedule:
Monday - Rest Tuesday - VO2 Max & Anaerobic Capacity Wednesday - Sweet Spot intervals Thursday - Recovery Ride Friday - VO2 Max & Anaerobic Capacity Saturday - Long Ride (3-4 hours) Sunday - Rest
It gave 6x5@100-105%, 4x15 SS and 3h steady state with no rest week.
Next prompt:
Review this plan and suggest improvements.
Weeks 1-2 & 3-4: Adjust long ride intensity
To be fair that didn’t really need to be adjusted but also it wasn’t a complicated plan to ask it to perform.
I’ve been experimenting with a GPT coach for about a month now and the benefits were outweighing the negatives right up until it started to what I assume is running out of memory.
I run it in a dedicated project within ChatGPT using 4o model. It was going super strong and I was getting used to immediate feedback pretty much mid-ride (I could quickly send a message saying I completed a planned climb in x:xx time together with my perceived effort and it would give me fueling adjustments, tips for how to complete the route and more—feels like a superpower or having a support car with a coach inside). All this on top of reviewing rides (and adjusting the weekly plan accordingly), dialing in food and hydration pre-, during- and post-ride (great results there) and general cycling banter with my personal context in mind.
It also generates .zwo workout files for all indoor rides which is super cool and seamless. It would describe a workout, we’d adjust a few things and then it spits out a file I upload to Rouvy and away I go. Super handy.
Unfortunately in the past few days it started to hallucinate more (claiming rides I’ve never done or misestimating climbs based on data it made up) and forget things like yesterdays ride. This quickly demolished whatever trust I had going for it and was a pretty sobering reminder that this is not a talkative coach but rather an LLM doing its best at putting sentences together.
I asked it to compile a single message describing everything it knows about my cycling hobby and progress so far so that I could use it as a base prompt for Claude AI that I want to try next.
I was a click away from puling the trigger on TR subscription, but I’m not going to lie—the feature allowing to talk to ai about your rides and discuss in plain words what you could’ve done better etc. is absolutely life changing and I hope it finds its way to TR soon—truly liberating and feels like future.
1 Like
I’m glad this is working for you, but I’m not gonna lie, this sounds absolutely exhausting to me 
its like instead of being in the moment and just paying attention to your body and surroundings, you’re instead focused on your cellphone and the minutiae of training.
I do like the zwo file creation idea. I’ll have to play around with that.
4 Likes
I think once you play around with the Intervals.icu MCP server for Claude (Anthropic’s LLM) you can definitely see that “one on one” coaching using AI agents is coming and it will be probably be really, really good. It will be interesting to see how TR can compete.
chatgpt maybe not. But there’s already quite a couple of AI gen or AI assisted coaches out there where it does work well.
I’ve been playing with the Intervals.ICU MCP server for Claude aswell (based on someone elses feedback on the forum).
Like @agapov describes, it’s really nice to be able to chat about your activities and get the AI to analyse your data.
At the moment, i’m still using the free Claude plan, so the total length of the conversation is always limited to a single activity before you get to the limits of the free plan. But I have played with the idea of subscribing to Claude and let it make my schedule.
@agapov What LLM are you using and running into this issue? Is this also Claude with the Intervals.icu MCP, or are you using something else?
Hahah, sorry I didn’t specify—I would check with Coach GPT when at a stop or a regroup when there’s time and I’m bored lol, definitely not pulling out a phone mid-ride to check if my cadence is on point hahahah.
1 Like
I’m using a basic paid ChatGPT plan where I created a Project so that all cycling chats and context are in one “place”. Not sure how much it helped, since it seems to still “forget” things if they’re discussed in different chat threads. I’m also pretty talkative and ask a lot of questions, so that might be a contributing factor.
I’ve noticed if I start every thread with a bit of a background prompt it helps the LLM orient itself in time better. Something like “Last week I did 8 hours of training over 105 miles and 9K ft of climbing. Sunday: rest, Monday: indoors, 60 minutes of Z2 with high 3x60s cadence spin outs, Tuesday: indoors, 4x15m of SS stacks at 86%, Wednesday: rest” etc. I would then try to ask as pointed question as possible, like “Build me a training schedule for this week knowing I’m planning a 100-mile group ride with 26ft of climbing for Saturday, as well as keeping focus on my goal of climbing Mortirolo in under 20 minutes”.
Specific segment goals are also super handy—I would often use GPT’s help with pacing strategy when doing a big effort or when a route involves various climbs at various points—extremely helpful since it pretty much gives me my power targets for each segment and explains why it chose those. The ability to ask follow up questions is absolutely golden, since I often bring in very personal context to such planning like having a birthday dinner before a ride or maybe running on low sleep—Coach GPT adjusts the plan and gives megabrain tips that I blindly follow assuming those are legitimate hahah.
I’m curious about diving deeper and building a dedicated MCP server for just my cycling struggles—currently its memory loss is very annoying and ruins a lot of the fun. Technology can’t advance fast enough for us guys!
Just use the Intervals.icu MCP server with Claude. You can get it to review your actual historical training data, tell it your goals and get it to generate a plan and create workouts. People have been using prompts like “act like a mix of Joe Friel, Dr. Stephen Seiler, and Dr. Andrew Coggan”.