Athmex

Cycling training with ChatGPT: the numbers it needs from you

Give ChatGPT your FTP, your real weekly hours, which days can take intensity, your event and its date, and anything that limits you. The catch is that cycling prescriptions are generally percentages of FTP, so a stale or estimated FTP quietly rescales the whole plan: sweet spot written from an old 308 W lands at threshold if you are actually 285 W.

What a usable cycling prompt actually contains

Ask ChatGPT or Claude for a cycling block and you will find the same prompt shape recommended everywhere, usually something close to: a four-week block for a 48-year-old cyclist, FTP 308 W, seven to nine hours a week, intensity Tuesday and Saturday, long Zone 2 Sunday. That shape is correct. It works because it carries the five things a plan cannot be written without: your FTP, the hours you actually have, which days you can go hard, the event you are pointing at, and anything that limits you.

The advice usually stops there. What almost nobody covers is what happens when one of those numbers is wrong, and the FTP is the one that goes wrong most often.

Your FTP decides everything after it

Cycling prescriptions are written as percentages of FTP, so the number is not one input among five. It is the scale every other session is drawn against. Get it wrong and the plan is still internally consistent, still looks professional, and is wrong all the way down.

Sweet spot is commonly defined as roughly 88 to 94% of FTP. Tell an assistant your FTP is 308 W and it will prescribe sweet spot at about 271 to 290 W. If your FTP has actually drifted to 285 W, those same watts are 95 to 102% of your real threshold. The session written as sweet spot is now threshold work or slightly above it, and if the block has three of them a week, you are riding over threshold three times a week while reading the word sweet spot on the page.

Nothing in the conversation will flag this. The assistant has no way to know the number aged, because you typed it once, months ago, into a different chat.

An estimated FTP is not a tested FTP

The second failure is harder to spot. If you connect a training platform, the assistant may be able to read an FTP from it, and that value may never have come from a test. Strava's connector exposes heart-rate zones, power zones and FTP, and it carries a flag recording whether the FTP was estimated rather than set by you.

That flag matters and it is easy to lose. An estimated FTP is a platform's inference from your recent rides. It is a reasonable starting point and it is not the same claim as a number you produced in a twenty-minute test, or a ramp test, or a hard forty-minute effort you know the shape of. If the assistant reads the value without reading the flag, an estimate becomes a fact, and every percentage after it inherits that.

Neither problem is the assistant being careless. Both are a missing date and a missing provenance on a single number, which is a storage problem rather than a prompting one, and the specific gap Athmex was built to close.

Same athlete, same request, one number different
What is being comparedPlan built on a stale 308 WPlan built on a current 285 W
Sweet-spot target (88 to 94% FTP)271 to 290 W, which is 95 to 102% of the athlete's real threshold251 to 268 W, which is genuinely below threshold
What the session actually isThreshold work, or just over it, labelled as sweet spotSweet spot, as written
Three of these in a weekThree threshold sessions the plan did not intend to prescribeThe intended load
How it shows upSessions that keep failing in the last interval, read as poor formSessions that finish as designed

Heart rate drifts too, with less noise

Threshold heart rate and maximum heart rate have the same property and get revisited less often. Heart-rate zones are derived from them, so a threshold heart rate set two seasons ago quietly moves every zone boundary that depends on it. The number changes with fitness, and it changes with age, illness, heat and altitude in ways that make a single old figure a poor anchor.

An assistant can derive zones, and it should, but only from an anchor value. The zones are not the thing to keep current. The anchor is.

The numbers worth pinning down

For a cyclist, the short list is genuinely short:

  • FTP in watts, with the date it was set and how it was set.
  • Threshold heart rate and maximum heart rate, with their dates.
  • Weekly hours you actually have, not the ones you would like.
  • Which days can carry intensity, which is usually a life constraint rather than a training one.
  • The event you are pointing at, its date, and whether it is a target or a training day.
  • Anything that changes what you can do: an injury, a work block, a knee that objects to standing efforts.

Everything else ChatGPT or Claude can reason its way to. These six it cannot, and it will not tell you they are missing. It will simply answer as though it knew.

Where Athmex fits

This is the part Athmex exists for. You enter those values once, each stored with the date it actually applied, and connect Athmex to Claude or ChatGPT directly. FTP is stored in watts, threshold and maximum heart rate in beats per minute, alongside your weekly hours, available days, constraints, goals and races.

Two details matter more than the storage. When the assistant reads your context it also receives how old each value is, and it is told to say so rather than quietly reason from a stale figure, which is the difference between a plan built on 308 W and a plan that opens by asking whether 308 W is still true. And because previous values are kept rather than overwritten, a threshold from March is still visible next to the one from July, which is what makes progression a question you can ask.

Athmex stores the anchors, not the zones, and it stores no rides at all. Your training history stays with your training platform and reaches the assistant through that platform's own connector. Athmex is the layer that says which athlete those rides belong to.

Common questions

Can ChatGPT see my FTP automatically?

Not on its own. It can read an FTP if you connect a platform that exposes one, and on Strava's connector that value may be an estimate rather than a test result. Otherwise the only FTP it has is the one you typed into that conversation.

How often should I update my FTP?

That is a training question rather than a software one, and it depends on how you test and how your block is structured. The useful software answer is narrower: whenever it changes, record the date it changed, because a number without a date cannot be judged for freshness by anything reading it later.

Should I give it my power zones or my FTP?

FTP, with its date. Zones are derived from it, so an assistant given the anchor can produce zones on whatever model you prefer, while an assistant given only zones cannot work backwards reliably or notice when they went out of date.

Does Athmex analyse my rides or write my plan?

No. Athmex stores no activity history and does no coaching. It supplies your confirmed numbers to the assistant you chose, and the assistant does the reasoning.

Cycling training with ChatGPT: the numbers it needs | Athmex