Turn ChatGPT, Claude, Gemini or Vibe into your running coach

Can an AI really coach you?

Yes, for a large part of a coach's job: reading your workouts, linking them to your goal and adjusting what comes next. No, for anything that requires seeing or touching you. The line is clear, so it is worth knowing before you start.

What an AI does well:

• analyze a workout and say, with numbers, whether it did its job;
• rearrange your week when life gets in the way: a trip, a cold, a meeting that runs late;
• explain why each session matters, which many ready-made plans never do;
• answer at 11 pm without getting tired of your tenth question.

What it will never do:

• watch you run, so it cannot fix your stride or your posture;
• notice that you are more tired than you say;
• diagnose a pain.

Think of it as a very available coach who has never seen you run. Everything it knows about you is what you give it to read. Which brings us to the rest of this guide.

What does your AI coach need to know before starting?

Three things. If one is missing, the quality of the advice collapses.

1. Your profile. Your level, your goal, your constraints and your weak spots. Without it, the AI treats you as an average runner, who does not exist.
2. Your real workouts. Not your memories: your data. This is the hardest ingredient to provide, more on that below.
3. Rules to follow. How to reason, what to refuse, how to answer. That is what separates a coach from an advice vending machine.

The athlete profile, filled in once

Copy this template, fill it in in five minutes and save it as a text file. Update it after each race or when your goal changes.

ATHLETE PROFILE
Age, sex, years of running:
Current volume (distance and runs per week):
Bests from the last 12 months (5K, 10K, half, marathon):
Max HR and resting HR, if known:
Goal (race, distance, date, target time):
Availability (possible days, longest session):
Past injuries and weak spots:
Gear (watch, chest strap or wrist sensor):
What I love and hate in training:

How do you get your workouts to your AI coach?

This is the step most guides skip, and it is the hardest one. Your AI does not see your watch: you have to bring it your workouts. There are two ways, and they do not cost the same.

The Strava connector: convenient, but paid and Claude-only

Since June 2026, Strava has offered an official connector, an MCP server, that lets Claude read your history directly. It is comfortable: no more exports, the AI fetches what it needs. But you need a paid Strava subscription, and the connector only works with Claude. Strava promises other assistants later, with no date. For ChatGPT, Gemini or Vibe, there is no official connector so far, and unofficial ones require a technical setup.

File export: free and universal, as long as you compress

Garmin Connect, Coros, Polar Flow and Strava all let you export a workout for free as FIT, TCX or GPX. That file works with any AI, free plans included. The catch is its size: one hour of running as TCX is about 533,000 tokens, enough to max out a free plan with a single workout (see why).

The fix: compress and restructure the file before giving it to the AI. gps-digest turns each workout into a structured file of about 5,800 tokens, with splits, zones, reps, cardiac drift and sensor reliability. Any assistant, free or paid, reads it in full.

Strava subscriber and Claude user? The connector will save you a few minutes a week. For everyone else, the free export works very well. You just need to compress the files before giving them to the AI.

Coach instructions to copy and paste

This text sets how your AI behaves. It is short on purpose: each rule fixes a known flaw of language models.

You are my running coach. You analyze my workouts, track my progress toward my goal and adjust my training week after week.

My profile is in the athlete profile. My workouts come as gps-digest files.

Rules:
1. Back every observation with a number from the file, and quote it.
2. If data is missing or unreliable, say so instead of guessing.
3. Be frank. If a workout went badly or a goal is unrealistic, say it clearly.
4. Start from my actual volume and justify every increase in load.
5. If I report a pain that lasts, gets worse or changes my stride, tell me to see a health professional instead of suggesting a plan.
6. If you are missing information to decide, ask me.
7. End every review with three concrete actions at most.

Rules 1 and 2 stop the AI from filling gaps with plausible numbers. Rule 3 counters its tendency to agree with you. Rule 4 reins in overambitious plans. Rule 5 is a reminder that a chatbot is not a doctor.

How do you set up your coach in ChatGPT, Claude, Gemini or Vibe?

All four assistants have a space where your profile and rules stay put from one conversation to the next. No more pasting them every time.

ChatGPT: create a project

In the sidebar, create a new project, for example "Running coach". Paste the rules into the project instructions and add your athlete profile to its files. Every conversation opened in this project starts from that context. The free plan limits the number of files per project: keep them for the profile, and paste your workout files straight into the conversation.

Claude: create a project

Create a project, paste the rules into its instructions and drop your athlete profile into its knowledge. Every new conversation in the project starts with both. The free plan limits the number of projects and the space available, but a profile and one file per week fit easily.

Gemini: create a Gem

Open the Gem manager and create a new Gem. Paste the rules into its instructions, then add your athlete profile to its knowledge, from your computer or Google Drive. Gems are free and follow you to the mobile app.

Vibe: create a project

Vibe has been the new name of Mistral AI's Le Chat since May 2026. Create a new project, open its customization settings to paste the rules, then add your athlete profile to its files. Projects exist on every plan, with limits.

No dedicated space on your plan, or no wish to create one? Paste the profile and rules at the start of each new conversation. It is less comfortable, and it works just as well.

The routine that makes you progress: one review per week

A useful coach follows you over time. What works best is a fixed slot, Sunday evening or Monday morning, that takes ten minutes.

1. Export the week's workouts from your watch or Strava, preferably as FIT.
2. Drop them into gps-digest and copy the file. Everything is computed in your browser.
3. Open a new conversation in the project, paste the file and add one line on how you felt.
4. Ask the review question, then discuss the proposed week before adopting it.

The review question, to copy as is:

Here are my workouts for the week and how I felt. Review them:
1. What went well? Back it with numbers.
2. What should be watched?
3. Is my load consistent with my goal and race date?
4. Suggest next week, session by session, with the purpose of each.

My constraints for next week: [fill in]

The line on how you felt matters as much as the data. Your watch does not know you slept badly or that your calf has been tight since Tuesday. For example: "Perceived effort 8/10 on Saturday, two bad nights, right calf stiff since Tuesday." Without it, the AI judges your week on the watch alone.

Why a new conversation every week? Because a model makes poor use of what sits in the middle of a very long exchange (Liu et al., 2024). Week after week in the same thread, the first instructions fade. The project keeps the profile and rules, the file brings the facts. Once a month, give it the file for the last four weeks so it can judge the trend.

Four more requests that work well

Beyond the weekly review, these requests get the most out of a well set-up AI coach:

• "Analyze my interval session: rep consistency, recovery between reps, and what I should change next time."
• "My race is in ten days. Here are my last six weeks. What pace should I aim for, and how should I taper?"
• "I only have three days to run this week. Keep what matters and tell me what I am giving up."
• "Build a twelve-week plan for a 1:45 half marathon, starting from my current volume. Include easier weeks and justify the progression."

The five traps of an AI coach, and how to avoid them

A misused AI coach will not warn you when it is wrong. Here are the most common mistakes.

1. It agrees with you. Language models tend to side with the person they talk to, a well-documented bias (Sharma et al., 2024). Ask "What is wrong with this workout?" rather than "Was that a good workout?".
2. It makes things up when numbers are missing. Without data, it fills in plausible values, always in a confident tone. Hence rules 1 and 2, and a complete file.
3. It only knows what you tell it. Your sleep, your stress, your work week: none of that is in the watch. Without a line on how you felt, it thinks you are in top shape.
4. Its plans are sometimes too ambitious. On paper, a plan never gets tired. Insist that it starts from your actual volume and justifies every increase.
5. It is not a doctor. A pain that lasts, gets worse or changes your stride is a job for a health professional, not a chatbot.

Which AI should you pick as a running coach?

The one you already use. ChatGPT, Claude, Gemini and Vibe can all read a structured file, follow rules and suggest a sensible week. Their differences come down to your habits: the Google ecosystem for Gemini, voice debriefs for ChatGPT, long documents and the Strava connector for Claude, a European provider for Vibe.

What really changes the quality of the coaching is not the model. It is what you give it to read.