STAS
Guide

Training with an AI chat: where it's strong and where it goes wrong

ChatGPT and Claude can review a season, build a week and explain any metric in a minute. They also average, agree with you, confidently fill in what they don't know, and never ask first. Know this in advance and the AI chat won't disappoint you: you'll ask in a way that gets you something useful, and check where it goes wrong.

In short: an AI chat is a helper, not a coach. Its answer isn't trusted by default; it's checked. The decision, and the responsibility for the result, are yours.

Where an AI chat is strong

All of this works when it has your data. Without it, it's advising an average athlete.

Gets through a lot of data fast

Seasons of workouts, heart rate, pace, power, your reports: a minute instead of an evening with spreadsheets, and it boils a week or a season down to a short summary. "Review the last two months: what improved, what slipped?"

Notices what you missed

A dip after every hard week, heart rate creeping up at the same pace, a long run that hasn't grown in three months.

Explains terms and metrics

What fitness and fatigue mean, what a threshold is for, why your heart rate on easy runs is higher than usual: no reference books or forums, just your own numbers.

Explains why a workout is in the plan

It knows the general principles of training and can say what each session develops. If it can't explain one, that session has no place in the week.

Rebuilds a week in a minute

A flight on Wednesday, a bad night's sleep, a cold: it reshuffles the week as soon as you tell it what's changed.

Pulls scattered pieces into a plan

Your coach's notes, an article, a race date, your work schedule and three days away on business: it turns them into a clear week where every session has a place and a reason.

Compares options quickly

Move the long run or skip it, two hard sessions or one, race in four weeks or in eight: it lays out the pros and cons of each so the decision is easier.

Helps you prepare for a race

Pacing for the distance from your own paces, tactics for the terrain, the general principles of fuelling and tapering. Gels, shoes and the pre-race breakfast: only what you've already tested in training.

Helps you shape a goal

It turns "I want to be faster" into a goal with a date and checkpoints, and asks what matters more: the time, the place or just finishing. From the goal comes the strategy.

Doesn't tire and doesn't judge

You can ask the same thing ten times, admit to a missed week or to not understanding a term: the tone won't change. Keep asking until you understand what the answer is built on.

Where it goes wrong and what to do about it

An AI chat's mistakes are predictable, and there's a fix for each. Below are the traits worth knowing before your first plan: how it handles data, what it doesn't know, how it answers and what it isn't accountable for.

  1. It averages

    Without your data it advises an average athlete, and even with your data it drifts towards a standard plan: three hard sessions a week, a long run from a table. Ask which of your workouts a number is built on.

    What people usually write

    Make me a half marathon training plan.

    Better

    Build a week from my history. Which of my workouts from the last month are the pace and the long run based on?

  2. It overloads the plan

    It reads "improve the plan" as "add to it": more intervals, more strength work, more kilometres. It can take things out, but only if you ask directly.

    What people usually write

    What else should I add to progress faster?

    Better

    What's unnecessary in this week? Give me a version with the same effect but one hard session fewer.

  3. It mixes up numbers, dates and units

    Kilometre totals, percentage increases, days of the week and the number of weeks to your race: recount them yourself or ask it to show the calculation step by step. Name units explicitly: it can mix miles with kilometres, pace per kilometre with speed, heart rate with power.

    What people usually write

    Volume is up 15%, good.

    Better

    Show the calculation in kilometres and min/km: how much last week, how much this week, and how many full weeks until April 12.

  4. It knows how you feel only from your words

    Sleep, pain, fatigue and mood: it knows them only from what you write or from your reports. Start with what's changed and give facts: where it hurts, since when, after what.

    What people usually write

    My knee is a bit sore, what do I do with the plan?

    Better

    My knee has hurt for three days since the long run, especially uphill. I'm seeing a doctor; until then, build a week with no downhill running or jumps.

  5. It can't see technique or conditions

    It won't see how you run, and heat, altitude, wind, hills and the surface don't exist for it until you write them down. It will put a high heart rate at 30°C down to fatigue if it doesn't know about the heat.

    What people usually write

    Why was my heart rate so high on an easy run?

    Better

    Easy run at 31°C, hilly, the day after a flight. Heart rate higher than usual: which of those explains the difference, and what's worth checking?

  6. It remembers only the current conversation

    It doesn't remember past conversations, or only partly, and in a long chat with a lot of data it loses details, fast models sooner than strong ones: a limit you named an hour ago can drop out of the plan. Keep what matters outside the conversation, start a new chat for a new task, and before saving ask it to check the plan against your rules.

    What people usually write

    Remember I mentioned my knee earlier?

    Better

    Before saving, check the plan against my rules: knee, no downhill running or jumps. What does it break?

  7. It takes no initiative

    It won't remind you about a workout, ask how the long run went or notice that you haven't written for three weeks. You start the conversation, and what it didn't ask, it doesn't know. Ask it to put its questions to you before the plan.

    What people usually write

    Build next week.

    Better

    Before you build the week, ask me about everything you're missing: how I feel, how much time I have, what's changed.

  8. It follows your lead

    It agrees with whatever the question hints at and believes what you've told it about yourself, even when the data says otherwise. Ask neutrally and give facts instead of conclusions.

    What people usually write

    I've recovered, put my old volume back.

    Better

    I was sick for a week, no fever for three days. Can I go back to my usual volume? What are the pros and cons?

  9. It sounds confident even when it's wrong

    It fills gaps in the data with a guess and doesn't warn you. Let it say "I don't know" and ask it to flag what it's unsure about.

    What people usually write

    Give me an exact half marathon prediction.

    Better

    How sure are you about the prediction? What in it is a guess, and what's missing to be more certain?

  10. It invents sources and studies

    Asked to back up its advice, it may name a study, an author or a figure "from the science" that doesn't exist, or credit a method with something that was never in it. Ask it to base things on your workouts, not on the literature, and check any references yourself.

    What people usually write

    Prove this is scientifically sound.

    Better

    Which of my workouts is this based on? If you cite a study, name it exactly and I'll check.

  11. The answer depends on the wording

    The same question asked differently, or tomorrow, can produce a different plan. That's not a reason to trust the first answer; it's a reason to compare: where the versions differ, the plan rests on a guess.

    What people usually write

    Got it, saving the first version.

    Better

    Give me two versions of the week, tell me the risk of each and why you'd pick one of them.

  12. Its knowledge isn't always current

    It learned from texts written over the years, some of them debatable methods and outdated norms, and it won't tell you which one you're looking at. If its advice differs from what your coach or doctor says, don't assume the chat is more up to date: ask which approach it's relying on and what the alternatives are.

  13. It isn't accountable for the result

    It won't find out how the week it suggested turned out, and it loses nothing if it's wrong. So the final say is yours, and checking the plan before it's saved is your job, not its.

When you need a person

A coach

Sees your technique and sees you: fatigue in the way you walk, nerves before a race, the things you'd never write down. Takes the decisions and answers for them. An AI chat doesn't replace a coach; it lightens the load: data analysis, drafts, questions between sessions.

A doctor or physiotherapist

Pain, an injury and the return from it, feeling strange: an AI chat can't examine you or prescribe treatment; it can only adapt the plan to what the doctor said. What belongs to the doctor rather than the plan is listed below.

How to get the most out of it

  1. The strongest model, not the fastest

    There's a lot of data: seasons of workouts, reports, thresholds. A fast model loses details and gives you generalities. For reviewing your history, your strategy and checking a plan, pick the strongest model available in your ChatGPT or Claude. You switch models inside the chat itself.

  2. A new conversation for a new task

    A long chat is the main source of lost details. Review the season, plan the week and ask about the race in separate conversations. Through STAS your data is visible in every new chat, and your goals, rules and strategy live in your Profile, so starting fresh costs nothing.

  3. Ground rules in the first message

    It follows rules set in the chat itself better than a wish to "be a coach". Three lines at the start change every answer below them: ask for what's missing, show the basis, flag uncertainty, and suggest what to cut, not only what to add. You can save them in the settings of ChatGPT and Claude so you don't have to repeat them.

    What people usually write

    Be my coach.

    Better

    Here's how we work: before a plan, ask me what you're missing; for every number, name my workouts and dates; flag what you're unsure about; tell me what to cut, not only what to add.

  4. "Why this?" and "what's missing?"

    This works better than "double-check yourself": without new data the model just repeats its answer. Asking for the basis makes it name workouts and dates, and asking what's missing makes it admit the gaps.

    What people usually write

    Double-check yourself.

    Better

    Which workouts and dates is this based on? What was missing?

  5. A second opinion from another model

    Ask another model the same question, for example Claude after ChatGPT: through STAS both see the same data. Where their conclusions differ, ask each one why it decided that way and what it's missing. Agreement isn't proof, but a disagreement shows exactly where the plan rests on a guess.

    What people usually write

    The other model said something different, redo it.

    Better

    The other model puts the long run at 18 km, you say 22. Which of my workouts is your version based on, and what are you missing to decide?

How it works in STAS

Your data in every new chat

Through STAS, ChatGPT and Claude read your workouts from Intervals.icu, your calendar, reports, thresholds and estimated paces. There's no need to retell your history in every conversation, and your own numbers replace the "average athlete".

Profile, rules and strategy outside the conversation

Goals, limits, rules and strategy are kept in your Profile, and the AI chat reads them at the start of every conversation. The knee you mentioned an hour ago won't drop out of the plan: it's written down.

Preview before saving

A plan or a change to your profile, goals or rules is shown to you first and saved only after you say yes. It's easier to check before it's saved.

Undoing profile changes

If the AI chat has saved the wrong thing to your profile, goals or rules, ask it to show the change history and bring back the previous text. It shows what it will restore and does it after you confirm.

What STAS doesn't change

The AI chat still averages, agrees, overloads the plan and gets the arithmetic wrong, and it still knows how you feel and what the conditions were only from your words and reports. STAS gives it your data; asking and checking is still up to you.

Health questions are for a doctor

Pain that doesn't go away, a fever, dizziness, an irregular heartbeat: these are questions for a doctor, not for an AI chat or STAS. The plan adapts to how you are, but it doesn't diagnose anything.

Common questions

If an AI chat gets so much wrong, what's the point?

It takes on the data analysis, the draft plan and the explanation of the numbers: the things that otherwise eat an evening. Its mistakes are predictable, and there's a fix for each: facts instead of conclusions, asking for the basis, a second opinion. With them it's a strong helper; without them, a confident conversation partner.

Which AI chat is better: ChatGPT or Claude?

Through STAS both see the same data, and the rules for working with them are the same. What matters more is picking the strongest model in each, and asking both for serious decisions.

Do I need a coach if I have an AI chat?

If you have a coach, the AI chat doesn't replace them; it lightens their load: data analysis, drafts, questions between sessions. If you don't, an AI chat with your data gives you more than a plan off the internet, but it won't see your technique, give you an outside view or take responsibility for decisions.

What of my data does the AI chat see, and where does the conversation live?

Through STAS it reads your workouts from Intervals.icu, your calendar, reports, thresholds and your profile with goals and rules. The conversation itself is stored by ChatGPT or Claude under their rules, not in STAS. What you write there beyond your training is up to you: the plan doesn't need complete medical records or other people's data.

Can I trust the AI chat once it's connected to my data?

Data removes some of the mistakes: it stops advising an average athlete and remembers your rules. But it still agrees with the question and guesses confidently. You don't need trust by default; you need a check, and it takes a couple of minutes.

Try it with your own workouts

Connect STAS once and your AI chat sees your latest workouts, goals and calendar without you retelling them. Once you confirm a plan, it goes into your calendar.