Compare

You already train with ChatGPT. Keep the conversation. Change the reader.

Runners who train with ChatGPT, Claude or Gemini have the right instinct and the wrong reader. The instinct is thinking out loud with something that explains its reasoning and leaves the call to you. The reader waits for your prompt, reads what you paste instead of the file your watch wrote, forgets last month unless you upload it again, and quietly re-tweaks the week you agreed on. Tempo is that same conversation, built for runners.

What you are doing right

Runners who take their training to a general assistant keep something the running apps never gave them: the why. Ask why a session is there and you get an argument you can push back on. The praise runners give ChatGPT, in the threads where they argue about it, is that it goes into detail on why certain things are done.

Keep that. A runner who wants the reasoning before the instruction is doing the thinking a self-coached runner does, and doing it well. The question was never whether to think out loud with an AI. The question is what the AI is reading while you do.

Where the reader breaks

Nobody makes the first move

A general assistant waits for your prompt. The long run happened or did not, and nothing asks how it went. Runners who work this way describe the same routine: you remember to open the chat, you type the summary, and only then you get the answer. The noticing stays on you.

What runners do about it today: open the chat and type the week in, again, by hand.

The paste gets read. The file does not.

Your watch wrote a file: every stretch of pace and heart rate, cadence, ground contact, the fade in the closing kilometres. What reaches the chat is a screenshot or a summary line, so the reasoning stands on the summary. Since June 2026 Strava's terms close its data to any third-party AI, and Garmin offers no official connector, so the file mostly stays where it is. From your Garmin file, not a Strava summary: that is the whole difference between analysis and paraphrase.

What reaches the chat today: a screenshot, or a summary line.

The chat forgets last month, then drifts

Runners who train this way upload their week again and again, because the chat that agreed the plan last month does not carry it forward. Told not to change the agreed week, the model still tweaked pace, distance and heart-rate zones. A decision you made carefully last month is a blank this month.

What runners do about it today: upload the week again, and re-state the plan they already agreed.

A general model has the same confidence for everything

A general model hands over a good plan and a bad one with the same confidence, because nothing in it knows what is true about you. One runner asking for a marathon build was handed an opening long run of 18 miles. Ground truth does not live in the chat, only the last thing you typed.

The week does not hold still

Across real training conversations the same pattern keeps showing up: eleven schedule disruptions from nine different runners in four months. Work, travel, illness, low motivation. Nobody runs the plan exactly as written, and that is normal, not failure.

So whatever reads your training has to remember both what your watch recorded and what you said. Last month's niggle, last week's decision, the pace you said felt wrong. A reader that holds only the numbers will keep proposing the week you already declined.

Tempo, once

In practice: a proposal arrives with the data lines it came from, and you can disagree with it line by line. Last month's niggle and last week's decision reach today's proposal without you re-uploading anything. Cadence, ground contact and pace drift come out of the raw activity file. Tempo stays a sparringpartner, not a coach, and never becomes one. The decision stays yours.

Who this is for