Your personal AI cycling coach powered by your training data

Who it's for and why

If you:

• Train regularly on a bike with a head unit
• Want to follow a plan but aren't ready to pay $200-500/month for a personal coach
• Have tried canned plans from TrainingPeaks / TrainerRoad / Zwift and found they ignore your life (work, family, weather, illness, mood)
• Want an honest analysis of your training, not "great job, keep going"

— then this approach is for you.

What it is in two sentences

It's a combination of three simple things: a program that pulls your workouts from your head unit's service to your computer; an AI agent (Claude, ChatGPT and the like) that reads those files and discusses training with you like a real coach; and a program that turns the AI's plan into a ready structured workout you then see on your watch or bike computer.

The result is a personal coach who remembers your history, sees the real data from every ride, discusses plans by voice or text and adapts to your life.

Right now the data import/export is implemented for the Garmin ecosystem (Garmin Connect), but the approach is not tied to one vendor: the sync program (trainings_sync) lets you add new connectors to other services and head units, so the project is easy to extend to other platforms.

What you'll need

1. A computer (Mac, Windows or Linux) — it need not be powerful
2. A head unit and its service account with workout history (Garmin Connect is supported today)
3. A subscription to an AI agent that can work with local files: e.g. Claude Pro (€20/month) or ChatGPT Plus with similar features
4. 2-3 hours for the initial setup (guided by the AI itself)

No programming knowledge is required: the AI agent walks you through every install step.

How it works on a normal training day

Monday evening. You write: "There's a 40 km team time trial on Saturday. What's the plan?" The AI opens the folder with your recent workouts, looks at the actual data (power, heart rate, volume, recovery), remembers your history and proposes a 5-6 day plan: Tue Z2 1.5 h, Wed threshold 3x10, Thu rest, Fri short openers, Sat race.

Tuesday morning. "Review my workout" — the AI reads the workout file, shows average power per interval, analyses pacing and compares with targets.

Wednesday evening. "Didn't sleep well. Swap the intervals?" — the AI checks the week's load and answers realistically: "Move the thresholds to Friday; there's a risk to your immune system after that recent cold."

Thursday. "Rain is forecast" — the AI offers an alternative route, an indoor trainer session, or moves it to another day.

Friday morning. The AI builds a workout file for your head unit, with exact power zones and step descriptions right on the screen.

Saturday after the race. "Race was cancelled, I rode in the mountains instead." The AI reads the ride data, sees the real load, and immediately builds next week's plan around what you actually did.

One-time setup

This is a simplified outline — in practice the AI agent guides you. Don't be scared by the terms: usually it's a few lines pasted into a terminal window.

1. Workout sync program: https://github.com/dchernykh1984/trainings_sync — pulls workout files from your head unit's service into a folder and runs in the background. This is also where you add connectors to other services.
2. Workout generation program: https://github.com/dchernykh1984/training_plan_generator — turns a workout description into a ready structured workout for your head unit.
3. Connect the AI agent to your computer. For Claude this is the "Filesystem"/Desktop extension that grants access to the folders you choose.
4. Give the AI context about you. Fill in a one-time profile: age, weight, goals, target races, schedule, training principles, injury and illness history.

That's it. After that it's an ordinary conversation.

What makes this approach unique

1. Adaptability to real life

Canned plans are written 12 weeks in advance and don't know your throat hurts today, your daughter's birthday is Thursday, a storm is coming Wednesday, last Saturday you rode 5 hours instead of 2, you didn't sleep, and your FTP isn't what it was. The AI coach factors all of this in every conversation — it builds the plan around your week instead of squeezing you into a plan.

2. Honesty

Most apps motivate you ("Great workout!") — that's their business model. A good AI will say: "The data shows you hammered the last interval against a recovery plan — tomorrow will be hard," or "Your real FTP is likely lower than you think — here's the evidence from 6 workouts." It's a coach not afraid to disappoint you.

3. Memory of your history

The AI remembers all your workouts (files stored locally), months of conversations, your races, illnesses, recoveries, workout preferences and favourite spots for intervals. After 3 months it knows you better than a coach who sees a client once a week.

4. Pro-level data analysis

From a workout file the AI extracts exact power for any period, heart-rate and power zone distribution, cardiac drift, pedalling efficiency, comparison with history, and corrections for terrain, altitude and temperature. This is the level of WKO5 or TrainingPeaks Premium ($200/year).

5. Ready structured workouts on your head unit

You don't just get "do 4x8 minutes"; you get a ready structured workout that loads onto your watch/bike computer: the screen shows exact zones for each step, a timer and target cadence. Like a Zwift or TrainerRoad workout, but generated individually for you.

6. Voice discussion (optional)

Many AI agents support voice: after a ride you can sit in the car and just talk through how it went — no typing, no computer.

Cost

Total: about €20/month (the AI Pro subscription). A personal coach of similar quality costs €300-700/month.

Limitations and risks

Technical

• The AI can make calculation and rounding errors — don't take everything as absolute truth
• It sometimes "invents" data when it can't find something; asking "where did that number come from?" usually helps
• Analysis quality depends heavily on the model: weak models give weak analysis

Psychological

• The AI won't ask "where's your workout?" — you have to come to it
• It won't shout at you or pull you out of self-pity the way a human coach can. If you need that, it's a downside

Legal and medical

• The AI is not a doctor or a certified coach. Serious issues (injuries, heart conditions, overtraining) need a real specialist
• If you have medical contraindications to exercise, consult a doctor first

AI availability

• Subscriptions to the best AI agents can be unavailable or expensive in some countries
• Service quality changes over time, sometimes for the worse

Comparison with the alternatives

Where to start

1. Get a free Claude or ChatGPT Plus account for a month
2. Download a couple of workout files from your head unit's service (in Garmin Connect: right-click a workout, Export to FIT)
3. Upload them to the chat: "Please analyse this workout, I want to understand how I'm progressing"
4. Judge the answer. If you like it, move to a full automated setup; if not, try another agent or model

Conclusion

In 5 years all of this will be standard. Right now, in 2026, it takes a bit more technical effort than one would like — but it delivers the quality of personal coaching for the price of a Spotify subscription. If you're willing to spend 2-3 hours on setup, you get a 24/7 coach who remembers you, sees your data and adapts to your life.

Above all: the AI does not replace common sense and your body. If your body says "rest", rest, whatever the plan says. The AI is a very smart advisor, but the final decisions are always yours.

*Written in June 2026. Technology changes fast — for current setup details see the READMEs of the trainings_sync and training_plan_generator repositories.*