I Built 'AI Coaching Kuroko-chan' to Manage My Daily Training with Claude Code
Hello, I'm Hayachon!
This time, I'm going to share something a bit different from my usual weekly logs: a detailed look at the 'AI Coaching' system I've been practicing.
I tag every weekly log with 'AI Coaching,' but no one has ever asked me, 'What exactly are you doing?' I had some free time, so I decided to summarize it!
🤖 First of all, what is 'Kuroko-chan'?
'Kuroko-chan' is the nickname I gave to an AI agent tool called Claude Code.It's originally a programming tool, but I've been using it as my own personal training coach.
It all started when I watched a video by Ninnin-san on YouTube about 'AI Coaching' and decided to copy it. I started out just wanting to try it, but eventually, I started paying for it and running it on Claude Code. By the way, my journey with generative AI has been 'ChatGPT -> Gemini -> Claude AI -> Claude Code,' and I've finally settled on this current setup.
(｀・ω・´) {Wait, isn't there also coaching you get from actual people?
Some people might think that. Since I've never actually experienced that, I'm just speculating, but... I felt like, 'I'd feel bad if I didn't get results...' so I decided to try AI coaching, where the potential for damage is lower.
Well, that's how I started AI coaching.
Next, I'll talk about the routine I usually follow (^^♪
🌅 Morning Routine: '/ohayo' Command
Every morning, I just type '/ohayo' (good morning) to Claude Code, and my training for the day begins.
① Automatic Sleep Data Retrieval
It links with a training management service, so my sleep duration, score, HRV (heart rate variability), and resting heart rate are all retrieved automatically. I don't need to input anything myself.
② Only One Question About My Leg Condition
I only have to answer one question: 'How are your legs today?' using multiple-choice options. That's it.
③ Automatic Generation of the Day's Workout
Based on my sleep data, HRV, leg condition, and weekly fatigue (training load indicators like TSS, CTL, and ATL), the AI determines the workout I should do that day.
④ Automatic Setup in Zwift
The generated workout is automatically placed in the corresponding Zwift folder, so when I open Zwift, today's menu is ready to be selected immediately.
In other words, I don't spend any time after waking up worrying about 'which workout should I do today?' This is honestly the most appreciated part. By saving my brain's resources, I've been able to focus on the training itself.
🌙 After Training: '/End' Command
When you finish your workout, type '/end'.
It will ask for your RPE (Rate of Perceived Exertion, or simply 'how hard was it?'), and once you answer, that information is recorded as feedback, which serves as material for adjustments when planning similar workouts in the future.
📅 Weekly Review and Planning
On weekends, I use '/weekly review' to summarize the week's data and '/next week's plan' to have it set up the menu for the following week. It's surprisingly convenient that if I mention constraints like 'I have a business trip, so I can't ride on Tuesday,' it will restructure the plan accordingly.
And if I type '/weekly note', it even automatically generates a draft for my weekly note. In fact, this series of weekly notes is almost entirely created using this system.
📈 Changes Since Implementation
Looking at the numbers, here is the change between June 2025 (when I earned my Fuji Hill Climb Silver) and this year.
FTP: 282W → 325W (+43W)
PWR: 4.19 → 4.77 w/kg
Fuji Hill Climb: 69 minutes 23 seconds (Silver) → 64 minutes 19 seconds (Gold)
Of course, it's not just the power of AI; the accumulation of training itself and diet management (see my wife's note) are also significant, but I believe that'not having to worry about deciding on daily workout menus' and 'being able to manage fatigue levels with objective numbers'are definitely thanks to AI coaching.
🔧 Challenges and Points of Caution
To be honest, getting to this point was quite difficult.
At first, it was very hard to convey prompts carefully, and sometimes it wouldn't provide the workouts I expected or would give misguided advice. AI is not omnipotent, so there are times when you need to make a human judgment and stop it if you think, 'Isn't this strange?'
Also, what is subtly important is to properly convey your 'training background, weaknesses, and preferences' to the AI at the beginning.In my case, I have communicated traits like 'I'm more of a puncher, so I'm good at steep slopes but bad at maintaining a steady pace' and 'I tend to overpace under pressure,' so it returns advice based on that. Conversely, I feel that if this part is thin, the AI will only be able to say harmless generalities.
Conclusion
So, I have introduced the overall picture of 'AI coaching created with Claude Code'.
Honestly, I think there are many parts that can be started without specialized programming knowledge (I am not an engineer either). If anyone is interested, please ask questions in the comments. I will answer as much as I can!
I will continue to do my best with Kuroko-chan toward my next goal 🚴