How an AI coach actually builds your training program

The plan is the least interesting part of coaching.

That sounds wrong coming from someone who has spent two years building a platform that generates training plans. But it is the conclusion the work keeps forcing. A twelve-week program is a prediction, and predictions about human bodies degrade fast. You sleep badly for a week. A calf niggle shows up. Work explodes. By week four, the program written on day one is describing an athlete who no longer exists.

Good human coaches have always known this. They write a week, watch what happens, and rewrite. The plan is just the current best guess; the coaching is the loop. So when we built Pelaris, the question was never “can AI write a training plan”. It could, years ago, and so can a PDF. The question was whether AI could run the loop.

Here is how that actually works, end to end.

Step one: an intake that behaves like a first consult

Everything starts from a structured intake, the same territory a good coach covers in a first session:

• Goals, and the priority between them when there is more than one
• Sport and event dates, a marathon, a Hyrox, a footy season
• Training history and current numbers, honestly assessed
• Weekly availability, down to which days are blocked and why
• Equipment, environment, and injuries that constrain what can be prescribed

None of this is novel. What matters is what happens next: the intake is not used to pick the closest match from a template library. It is used to select and combine methodologies.

Step two: named methodologies, not invented ones

Pelaris does not improvise training theory. The generator draws from a library of named, established systems: 5/3/1, daily undulating periodisation, and block periodisation on the strength side, Pfitzinger, Hansons, Daniels, Maffetone, and 80/20 on the running side. The AI’s job is selection and adaptation, which methodology fits this athlete, this goal, this number of available hours, and how the pieces combine into one coherent week.

That last part is where most of the difficulty lives. Combining a strength system with a running system is not stapling two programs together. The interference effect is real, and managing it means the scheduler enforces rules a spreadsheet cannot: hard endurance and heavy lower-body work spaced apart, easy aerobic volume filling the gaps, deloads landing across both systems in the same week. Concurrent training is treated as a first-class problem because our users, hybrid athletes, team sport players, lifters who run, are exactly the people generic programs abandon.

Step three: the program is written in fortnights, not quarters

Pelaris generates the plan a rolling two weeks at a time. This is a deliberate design decision, not a limitation. Week six does not exist yet, and that is the point: when it is written, it will be written from the data of the athlete who just finished week five, not from a day-one prediction about them.

A twelve-week arc still exists, the periodisation, the build toward the race, the planned deloads. But the arc is a direction, and the sessions are computed close to the ground, where the information is freshest.

Step four: the loop, which is the actual product

Every session you log feeds the next decision. The workout tracker captures sets, reps, load, and RPE on strength work, pace, duration, and RPE on conditioning. A daily check-in takes about ten seconds and logs sleep, mood, soreness, and readiness. Individually these are small signals. As trends, they are the coaching:

• Squat RPE climbing across three weeks on the same loads, next week’s volume pulls back before the stall becomes a hole
• Readiness flat and sleep poor for a stretch, the deload comes forward rather than waiting for the calendar
• A tough match logged on the weekend, the following week reshapes around the recovery cost
• Sessions coming in easier than prescribed, loads progress faster

If you have read our piece on how an AI tracker breaks a plateau, this is the same principle running program-wide. The athlete does not have to notice the trend, run the analysis, and rewrite four weeks of programming. The system notices, proposes, and the athlete decides. What good AI coaching sounds like in conversation is a related but separate question; this loop is what it does regardless of whether you ever open the chat.

One line on privacy, because it shapes the architecture: Pelaris is built privacy-first. The coaching runs on training signals, your sets, your sleep score, your RPE, not on your identity. That constraint was set on day one and the whole pipeline respects it.

Where the recorded world ends and coaching begins

The fitness data ecosystem is extraordinary at recording. A Garmin on your wrist, a season of activities on Strava, a recovery score from a wearable, between them, a serious hobbyist athlete generates more physiological data than a professional had access to twenty years ago. I mean that without irony; the recording layer is a solved problem, and solved brilliantly.

What sits mostly unsolved is the layer above it: the system that reads what was recorded and changes what you do next. A watch can tell you last week’s training load was high. It cannot rewrite next week’s program around your half marathon, your Tuesday commitments, and the squat progression you are three weeks into. That gap between measurement and prescription is where Pelaris lives, and we built it device-agnostic on purpose: the coaching layer should sit on top of whatever records your training, not compete with it. I wrote previously about what general-purpose AI gets wrong about training plans; the short version is that reasoning without your data is advice, and data without reasoning is a diary. Coaching is the two connected.

What this looks like in practice

A concrete composite, the shape of a real user week:

A 38-year-old with a 90-minute lunch window on weekdays and long-run Sundays is eight weeks from a half marathon and wants to keep two lifting days. The generator pairs an 80/20 running structure with an upper/lower strength maintenance split, places the heavy lower session 72 hours clear of the Sunday long run, and holds Wednesday as the quality run. In week five, logged RPE on the tempo runs starts drifting up while sleep scores drift down. The next fortnight is generated with the tempo volume trimmed and the deload pulled forward a week. The race taper then builds off what was actually absorbed, not what was originally planned. Race week itself looks like a proper taper: volume down, intensity held, nothing new.

No single step in that story is magic. The value is that nobody had to sit down on a Sunday night and figure it out. That is the job the AI coaching layer does, week after week, and it is the job static programs, however well written, structurally cannot.

Try the loop, not the plan

The tracker is free forever, logging sets, RPE, and check-ins costs nothing, and even self-coached athletes get sharper the moment they start measuring honestly. The paid tier is the loop itself: the program that writes and rewrites your next fortnight from what you actually did. Start with how it works, or go straight in and start free.

If you are curious what honest effort scoring looks like first, the RPE guide is the right primer, and structuring your training week covers the scheduling rules the generator automates. The mechanism is not a secret. The advantage is running it every day without fail, which is the one thing software does better than any of us.