How the lope coach is built
A general-purpose AI doesn't know your training. lope's coach is a Claude model wired up to four sources: a coaching system prompt with a specific philosophy, four established methodologies as reference, a set of dedicated knowledge bases, and your full training context every time you ask a question.
The coaching philosophy
The coach has a system prompt that tells it how to behave: opinionated, evidence-based, data-driven, specific. It synthesizes four endurance methodologies into a coherent stance rather than picking one and parroting it.
Precise, purpose-driven workout prescription. Every run has a specific physiological target. VDOT-derived paces for all quality sessions.
The aerobic engine is paramount. Easy running should be genuinely easy. Toughness is psychological flexibility, not suffering. Lactate is fuel, not waste.
80/20 intensity distribution. Most runners run too hard on easy days. The single most consequential change most self-coached runners can make.
Marathon training demands race-specific work under fatigue. Long runs with marathon-pace segments. Training should be harder than the race in specific ways.
The synthesis matters. A coach that only knew Daniels would over-prescribe interval workouts; one that only knew Magness would skip enough quality work. The blend is the point.
The knowledge bases
Beyond the four methodologies, the coach has access to four dedicated knowledge corpora that sit alongside the system prompt every conversation. These were assembled from established sports-science literature and are deliberately bounded — the coach references them when relevant rather than improvising.
Training science
VDOT calibration, full periodization phases, gray-zone diagnostics, cardiac drift interpretation, recovery science, heat and altitude adaptation.
Injury prevention
The seven most common running injuries, return-to-running protocols, strength exercises for runners, when to stop / modify / continue. Always defers to medical evaluation for persistent pain — never prescribes injury treatment.
Race execution
Pacing strategy by distance, negative-split physiology, when-things-go-wrong protocols, taper psychology.
Sports nutrition
Daily fueling by training load, in-race carbohydrate targets, gut training, carb loading, hydration, RED-S awareness.
What the coach sees about you, every conversation
Each time you message the coach, lope builds a context block and prepends it to the conversation. This is what makes a generic AI into a coach that knows your training:
- Your last 10 activities, with all the per-run signals from per-run analysis — terrain-adjusted cardiac drift, route shape, fade detection, per-second zone distribution, plan vs actual, cadence, RPE, and the coach's own prior reactions to those activities
- 4-week aggregate stats — weekly mileage trend, intensity distribution, zone-2 pace progression, gray-zone time
- Your athlete profile — current goals, target races and dates, injury history, training preferences, lifestyle context
- Your last 10 chat messages in the current conversation, plus key statements from prior chats — so continuity holds across sessions
- Your current training scores with directional trend (improving / stable / declining) — see training scores
- Athlete notes — any freeform context you've added (HR strap glitching, hamstring tweak, sleep was rough this week)
All of it injected as text into the prompt every time. Generic chatbots can't replicate this without the data being structured the way ours is.
Athlete notes — the layer most apps don't have
You can leave the coach freeform notes about anything: an injury, equipment issues, life stress, sleep, fueling concerns, or a correction to the data (“HR strap was bad on Sunday's run”). These notes are factored into every coach response until you mark them resolved.
Practically, this means the coach won't flag an obviously-wrong HR reading as aerobic deficiency if you've flagged the sensor malfunction. It will adjust workout suggestions if you've mentioned a calf tightness. It's the layer that makes the difference between coaching that pretends your data is perfect and coaching that knows real life happens.
Continuity across chats
The coach remembers prior advice. If three weeks ago it told you to slow down on easy runs and you've been compliant, it acknowledges progress. If you've been ignoring a recommendation, it can gently revisit it. If a workout you're asking about now was discussed last week, the coach builds on what was said rather than restarting from zero.
This is implemented via a structured per-activity reaction history (the coach's prior comments on specific runs are stored alongside the activity) plus the recent message window from the current chat thread.
What the coach is not
Three places we've been deliberate about boundaries:
Not a doctor
The coach can discuss common running injuries and modification protocols. For persistent pain, sharp pain, or anything systemic, it always recommends professional medical evaluation. It will not prescribe treatment.
Not a 1-on-1 human coach for elites
A real coach who knows you intimately, watches you run in person, and adjusts based on observation in real time will likely outperform an AI for elite athletes. lope is built to be a strong upgrade over going it alone — not a replacement for human relationships at the top of the sport.
Not infallible
Treat the advice like a smart, well-read training partner. The coach is grounded in established methodology, but AI is fallible. Sense-check anything that conflicts with what you know about your own body, and ask follow-up questions if a recommendation surprises you.
Under the hood
The coach runs on Claude Sonnet via the Anthropic API. Each conversation streams responses token-by-token. The full system prompt, knowledge bases, and your training context are sent with every message — no fine-tuning, no embedding retrieval, just a deliberately structured prompt and a capable model.
Anthropic's API does not train on prompts or completions from API customers like us. Your training data isn't leaving lope's database for any purpose other than answering your specific question.