Methodology

The Training Science Behind Joules: How an AI Actually Coaches a Season

Joules June 28, 2026 · 10 min read

"AI cycling coach" sounds like a chatbot guessing. It shouldn't be, and good ones aren't. Here's the actual methodology underneath Joules — the periodization models, the intensity math, the load-to-recovery logic, and the honest places it stops — so you can judge it the way you'd judge any coach: on the reasoning, not the marketing.

If you've used a raw chatbot for training, you know the failure mode: confident, generic, and amnesiac. It restarts you every week instead of progressing you. The fix isn't a smarter model — the frontier models already know more training science than most riders. The fix is giving the model a coach's structure: an explicit knowledge base to reason from, a periodized plan to work against, real load math, and a feedback loop. That's what this post is about.

It plans from an explicit knowledge base, not from memory

Joules reasons against a structured cycling-science reference covering energy systems, adaptation timelines, periodization, intensity distribution, progressive overload, tapering, and discipline-specific prescription. That matters because it pins the coaching to known physiology rather than whatever the model happened to absorb. A few of the load-bearing ideas:

• All three energy systems are always on. Even 2–5 minute "anaerobic" efforts draw 60–80% of their energy aerobically, so the aerobic base gets trained to raise the ceiling at every duration — not just for long days.
• Intervals are designed around recovery, not just work. Sprint quality needs near-complete restoration (roughly a 1:8+ work-to-rest ratio); VO2max work uses 3–5 minute efforts with matching recovery to maximize time near maximal aerobic power; threshold work is capped around 40–60 minutes of time-in-zone before glycogen depletion erodes quality.
• Zones are individualized, not gospel. The fixed 2.0 / 4.0 mmol/L lactate landmarks are population averages; fiber type and genetics move them. When lab data isn't available, RPE and the talk test carry the load — which is also why Joules can coach without a power meter.

It chooses a periodization model — it doesn't have one default

The first real coaching decision is structural: how to shape the season so you arrive ready on the right day. Joules selects the model from your situation rather than forcing everyone through the same template:

• Classical linear — a single major event with 16+ weeks to prepare.
• Block periodization — well-trained or plateaued athletes, and stage races, where concentrated blocks (accumulation → transmutation → realization) drive a stalled system.
• Modified block — when you need multiple peaks across a season.
• Block with concentrated intensity — the time-crunched athlete under ~8 hours/week, where pure polarized volume isn't realistic.
• Reverse periodization — ultra-endurance events over four hours.

Phases then get real durations and intent: a base phase of 6–12 weeks at high volume and mostly easy intensity; a build phase that raises specificity and adds threshold/VO2max work; a short, sharp peak; and a taper. Training specificity increases as the event approaches — that progression is the whole point of a plan.

Intensity distribution: mostly easy, on purpose

How your weekly time is split across intensities is one of the highest-leverage choices in training, and the evidence favors keeping most of it genuinely easy. Joules works from the three standard distributions and matches them to your available hours:

• Polarized (~75–80% easy / minimal middle / 15–20% hard) for trained athletes chasing VO2max, when weekly volume supports it.
• Pyramidal (more easy, a band of tempo/threshold, a little hard) for base phases and high-volume blocks.
• Threshold / sweet-spot-leaning for the time-crunched athlete whose easy volume is simply too small for polarized to work.

The recurring mistake it's built to avoid is the "grey zone" trap — riding the moderate-hard middle so often that the easy days aren't easy and the hard days aren't hard. Zone 2 at the aerobic threshold is where mitochondrial density and lactate clearance get built, and it's protected accordingly.

The load math is the industry-standard kind

Coaching runs on measured stress, and the measurement has to be trustworthy. Joules computes Training Stress Score (TSS) using the Normalized Power algorithm — the same method behind TrainingPeaks and Intervals.icu — rather than a naive average, so variable-intensity sessions are scored accurately and line up with the numbers you already read elsewhere.

When the coach prescribes an indoor session, a workout solver builds the actual intervals to hit a target TSS within the right intensity band, using explicit zone bands and rest ratios and respecting how anaerobic work capacity (W′) depletes and reconstitutes during recovery. Every structured target is written as a percentage of FTP, so when your FTP moves, the whole plan recalibrates automatically. Don't have a current FTP? It can prescribe a ramp test first.

Progression and recovery are prescribed, not accidental

Getting harder is easy; getting harder sustainably is the craft. The rules Joules works within are the conservative, well-supported ones:

• 2–5% weekly load increase, raising volume or intensity but not both, and never exceeding a ~10% weekly jump.
• Load-to-recovery cycling — 3:1 (three weeks load, one easier) as the default, 4:1 for younger experienced athletes, and 2:1 for masters (50+), high-intensity phases, or anyone showing recurring fatigue.
• Deload weeks that actually deload — volume cut 41–60% while intensity is maintained, because intensity is what preserves the adaptations you've built.

Tapering follows the same evidence: drop volume 41–60% over roughly 8–14 days, hold intensity, and you can expect a performance bump on the order of a few percent. How fresh you arrive is tuned to the race's priority — sharper for an A-race, less so for a training-stimulus B-race.

The feedback loop is the part a static plan can't do

A plan you can't deviate from isn't coaching. Each Joules week is written with an explicit intent — build, recovery, peak, or taper — and a target load. Once your rides come in, the week is graded against what you actually did, and the next week is shaped by that result together with your fatigue-and-freshness trend (CTL / ATL / TSB) and your own post-ride reflections. Build weeks get harder on purpose; recovery weeks pull back on purpose. Nothing drifts.

Crucially, the load metrics are treated as instruments, not prescriptions. CTL, ATL and TSB point at directional risk — "you're digging a hole," "you're fresh enough to race" — but the literature is clear that individual response varies by 30% or more, so they inform the decision rather than dictate it. Your own history at a known freshness level beats a textbook number every time, and the coaching is built to defer to it.

The honest limits

A methodology piece that only lists strengths isn't honest. Here's where an AI coach — Joules included — genuinely stops:

• Estimates are estimates. Without power, load is modeled from duration and effort. That's enough to train well, but it's a model, not a measurement, and it inherits the honesty of your effort ratings.
• Population averages need individual calibration. Threshold landmarks, W′ values, and adaptation timelines are starting points. Real individualization comes from watching your power-to-HR-to-RPE relationship over weeks — which is exactly why the feedback loop matters.
• It is not your physiotherapist. AI can place intensity and recovery sensibly. It cannot diagnose an injury. Pain is a stop sign, not a data point to push through.
• You need a connector-capable assistant. On Claude that's any plan, including the free tier (free Claude allows one custom connector); on ChatGPT it's a paid plan (Plus/Pro) — that's the chatbot's rule, not ours. Joules itself is free while in beta.

The takeaway

The interesting question was never "can an AI talk about training?" — it always could. The question is whether it coaches from a defensible method: a periodization model chosen for your season, an intensity distribution that protects the easy days, load math the rest of the industry uses, conservative progression with real recovery, and a feedback loop that adapts to the rides you actually do. That's the bar Joules is built to clear — and to be candid about where it doesn't.

Want to see the method produce something concrete? Read a real Joules training week, day by day, or build a sample week yourself in your browser — no account or connector needed.