Your plan isn't decided by artificial intelligence.

It's code, rules written once and for all, and references we can name. This page tells you exactly what produces your sessions, and where their numbers come from.

Get notified

What the AI does, and what it doesn't

The question deserves a precise answer, because many apps simply say that artificial intelligence takes care of everything.

Code sets the calendar, the phases, the volume of each week, the structure of each session and its paces. None of this goes through a language model.

The AI writes each session's title, purpose and tip. A system instruction explicitly forbids it from writing a pace, a speed or a time.

If it writes one anyway, the text is automatically cleaned and replaced by a deterministic template.

The result: two people with the same profile get the same plan, and a plan doesn't change because a model was updated.

Measured benchmarks, discipline by discipline

Everything starts from an individual benchmark. Without it, a plan is just a copied template.

Running: maximal aerobic speed (MAS)

The one you know, otherwise estimated from a recent performance using Riegel's formula, otherwise a starting value based on your level. Paces follow from it: easy at 65%, marathon at 80%, threshold at 86%, MAS at 100%.

Swimming: critical swim speed

Derived from a 400 m time or an easy 100 m pace. Sets are then expressed in seconds per 100 m.

Cycling: speeds and heart rate

Reference speeds, and zones calculated using the Karvonen method, which takes your resting heart rate into account and not just your maximum.

What about training load?

It is calculated session by session using Banister's TRIMP model, with time constants of 42 days for fitness and 7 days for fatigue, the framework used in training physiology since the 1990s. It is used for tracking: it doesn't drive plan generation yet, and this page doesn't claim otherwise.

The rules that bound a plan

Paces are recalculated at the end of each block from the test session and the sessions actually completed, using a weighted median. A clearly abnormal test is discarded, and the change stays capped between minus 10% and plus 15%.

The coming week's volume moves by at most 10% either way, depending on how you feel and your form.

A recovery week every 4 weeks, and a taper at the end of the block.

In triathlon: never two hard sessions on the same day, at least 48 hours between two quality run sessions, the long ride never on the same day as the long run unless it's a brick, and one brick per week in the build and race-specific phases.

The references, named

The engine's settings don't come out of nowhere. Here is the work they're based on.

Periodization and week design: Joe Friel.

Running paces and long run proportion: Jack Daniels.

Predicting one performance from another: Riegel's formula.

Heart rate zones: the Karvonen method.

Critical swim speed: the work of Wakayoshi et al., 1992.

Intensity distribution: the 80/20 framework, and the work of Matt Fitzgerald.

Brick frequency: MyProCoach, Dave Scott, 220 Triathlon and TRI247, four independent sources that converge on one per week in the build and race-specific phases.

Training load: Banister's TRIMP model.

Questions about the method

Why not let AI do everything?

Because a training plan should be verifiable. When code decides, a rule can be reread, tested and corrected. When a model decides, all you can do is observe the result.

Can my paces be wrong?

They can be imprecise at the start if your benchmark is. That's why they are recalculated at the end of every block, from what you actually did and not just from a single test.

Does Trainetic replace a coach?

No. A coach sees what a calculation can't. Trainetic gives a consistent, quantified, adjusted plan to people who don't have a coach.