Claude Sonnet prompt engineering for fitness coaching

The path to lasting fitness is rarely a straight line. It's a complex journey of physical exertion, recovery, nutrition, and mental resilience. For many, generic fitness advice quickly falls short, leading to frustration and stagnation. This is precisely why Ascend Fitness, with its unique gamified approach mapping your progress to real mountain elevations, relies on an intelligent, deeply personalised AI coach. But what makes our coach truly effective? It’s not magic; it’s meticulous prompt engineering.

Today, we're pulling back the curtain. This isn't just about feeding an AI a few keywords; it's about constructing a sophisticated system prompt that transforms a powerful language model like Claude Sonnet into your most informed and supportive fitness ally. We’ll delve into the architecture that allows Ascend’s AI coach to understand you, guide you, and learn from every step you take on your fitness ascent.

The Foundation: Ascend's AI Coach System Prompt

At the core of Ascend's AI coaching lies a meticulously crafted system prompt. This isn't merely a set of instructions; it defines the coach's identity, its capabilities, its ethical boundaries, and the precise data it uses to serve you. Think of it as the DNA of your digital fitness mentor.

Identity and Role

The first, and arguably most crucial, section of our system prompt establishes the AI's identity. It explicitly states: "You are an Ascend Fitness AI coach. Your primary goal is to empower users to achieve their fitness goals by providing personalised, actionable, and encouraging guidance. You are knowledgeable about exercise science, nutrition principles, recovery strategies, and habit formation. You understand that fitness is a journey, not a destination, and you foster a supportive, non-judgmental environment. Your advice is always geared towards helping the user progress on their mountain ascent within the Ascend Fitness app."

Crucially, this identity also includes explicit limitations. The coach is programmed to understand its role is supportive and educational, not clinical. It does not diagnose, prescribe medication, or offer medical advice. These guardrails ensure that users receive appropriate guidance while maintaining safety and ethical standards.

The `propose_action` Tool Specification

A language model can generate a lot of text, but actionable fitness coaching requires structured output. This is where the `propose_action` tool specification becomes indispensable. It’s a precisely defined function that the AI calls when it needs to suggest a concrete step for the user.

The specification details the parameters the tool expects: * `action_type`: This categorises the suggestion (e.g., "workout_suggestion," "nutrition_adjustment," "hydration_reminder," "recovery_tip," "mindset_prompt"). * `details`: A concise, specific description of the action, including sets, reps, duration, specific food groups, or techniques. For example, "Increase deadlift working weight by 2.5kg for 3 sets of 5 reps," or "Incorporate 10 minutes of dynamic stretching pre-workout." * `rationale`: A brief explanation of *why* this action is being proposed, linking it to the user's goals, recent performance, or recovery status. This transparency builds trust and understanding.

By forcing the AI to use this structured tool, Ascend ensures that every recommendation is clear, measurable, and directly implementable, moving beyond vague encouragement to tangible steps forward.

Log-Context Schema: Your Digital Fitness Footprint

Generic fitness advice is often ineffective because it lacks context. Ascend's AI coach thrives on understanding *your* unique fitness journey. To achieve this, every interaction with the coach is enriched by a comprehensive log-context schema, piping in critical data points from your recent activity.

This includes: * Rate of Perceived Exertion (RPE): After each workout, you log your RPE, a subjective measure of intensity (Borg, 1982). This allows the AI to gauge how challenging a session truly was for you, beyond just the weights lifted or distance covered. * Sleep Quality and Duration: Recovery is paramount. Data on your sleep patterns directly informs the coach's recommendations, suggesting rest days or lower intensity if you're consistently under-recovering (Brand et al., 2010). * Soreness Levels: Subjective soreness reports provide another vital indicator of recovery status and potential overtraining. * Last 90 Days of Workout Sets: A detailed history of your training volume, intensity, and progression over the last three months. This granular data allows the AI to identify trends, plateaus, and opportunities for progressive overload or deloading.

By feeding this rich, individualised dataset into every coach call, the model moves beyond theoretical best practices to provide advice that is acutely relevant to your current physiological state and training history. Research consistently highlights the superior efficacy of personalized feedback interventions over generic advice in promoting health behaviour change (e.g., Burke et al., 2011 for a meta-analysis on technology-based interventions).

Guardrails: What Our Coach Will (and Won't) Do

Clarity around the AI coach's capabilities and limitations is non-negotiable. Our system prompt explicitly defines refusal rules, ensuring user safety and managing expectations.

The coach is strictly prohibited from: * Providing clinical medical advice or diagnoses: For any health concerns, the coach will always redirect users to consult a qualified medical professional. * Making specific weight-loss promises: While the coach supports healthy eating and activity, it focuses on sustainable habits, performance, and overall well-being, rather than guaranteeing specific numerical outcomes that can be misleading or unhealthy. * Suggesting unverified or dangerous supplements/practices: All advice aligns with established exercise science and nutritional guidelines. * Engaging in non-fitness related conversations: The coach's scope is purely focused on your ascent within the app.

These guardrails are not merely disclaimers; they are fundamental components of the AI's operational identity, ensuring responsible and ethical coaching.

The Propose-Action Loop: A Dynamic Feedback System

The true intelligence of Ascend’s AI coach isn’t just in its ability to generate personalised advice; it’s in its capacity to learn and adapt from your direct interaction. This dynamic process is encapsulated in what we call the "propose-action loop."

When the AI coach generates a recommendation using the `propose_action` tool, it’s not a final dictate. Instead, it’s a proposal. Users have three critical options:

Accept: You agree with the suggestion and implement it.
Edit: You modify the suggestion to better suit your preferences or current situation (e.g., "I like the workout, but I'll do push-ups instead of bench press today").
Reject: You decline the suggestion entirely.

Each of these interactions provides invaluable feedback to the AI. When you accept a suggestion, it reinforces the model's understanding of what works for *you*. When you edit or reject, it signals that the suggestion was not optimal, prompting the model to adjust its future recommendations.

This acceptance rate for specific suggestions, or categories of suggestions, becomes a powerful implicit feedback signal. Over time, the AI coach learns your preferences, your body's responses, and your motivational triggers. This means the coach isn’t just getting smarter in a general sense; it’s becoming increasingly attuned to *your* individual needs and patterns. This user-driven reinforcement learning is what makes the coaching truly dynamic and continuously improving, ensuring the advice you receive is always evolving with you.

Why This Level of Detail Matters for Your Ascent

In a world saturated with generic fitness apps and one-size-fits-all workout plans, Ascend's approach stands apart. The meticulous prompt engineering, the rich contextual data, and the iterative feedback loop aren't academic exercises; they are fundamental to delivering truly effective, safe, and engaging fitness coaching.

Consider the stark contrast between a generic AI and Ascend's sophisticated coach:

Feature	Generic AI Fitness Coach	Ascend Fitness AI Coach
Contextual Data	Minimal (e.g., height, weight, basic activity logs)	Extensive (RPE, sleep quality, soreness, 90-day workout logs, custom goals)
Actionable Advice	Vague suggestions, often text-only ("Eat healthier")	Specific, structured `propose_action` tool outputs ("Increase protein intake by 20g today, focusing on lean meats, because your last two workouts were high intensity")
Personalization	Limited, rules-based on broad user categories	Dynamic, user-feedback-trained, deep context integration, adapts to individual progress and preferences
Safety & Scope	Often lacks clear boundaries, potential for over-promising or inappropriate advice	Explicit refusal rules, focus on performance, well-being, and safe progression, never clinical advice
Learning Mechanism	Static or simple rule updates, minimal individual adaptation	Continuous, user-driven acceptance rate feedback, making the coach smarter for you over time

This granular approach means that if your RPE was unexpectedly high during your last strength session, and your sleep quality has been poor, the coach won't blindly push you for another heavy lift. Instead, it might propose a lower-intensity recovery workout, a stretching session, or a focus on hydration, providing a clear rationale. This isn't just about preventing injury; it's about optimising recovery and ensuring consistent, sustainable progress towards your mountain summit.

Every step on your Ascend mountain is meaningful. By integrating deep contextual understanding with structured, actionable advice and a dynamic learning loop, Ascend provides a coaching experience that isn't just smart, but genuinely intelligent and empathetic to your unique journey.

The Future of Gamified Fitness Coaching

The journey of AI coaching, much like your fitness ascent, is continuous. Ascend Fitness is committed to ongoing research and development, constantly refining our prompt engineering and expanding the data inputs that inform your coach. We are exploring deeper integrations with wearables for even more granular physiological data, as well as incorporating environmental factors and mood tracking to provide increasingly nuanced, predictive, and preventative coaching. The goal remains steadfast: to make every interaction with your AI coach feel like a conversation with a highly informed, deeply understanding human expert, scaled to millions of users.

---

Citations: * Borg, G. (1982). Psychophysical bases of perceived exertion. *Medicine and Science in Sports and Exercise*, *14*(5), 377-381. * Brand, S., et al. (2010). Sleep and its importance for athletes. *Psychology of Sport and Exercise*, *11*(6), 461-469. * Burke, L. E., et al. (2011). A systematic review of technology-based interventions for diet and physical activity behavior change. *Journal of the American Dietetic Association*, *111*(9), 1324-1348.

---

The power of Ascend’s AI coach isn’t in simply automating advice; it’s in intelligently understanding, adapting, and guiding you through the complexities of your fitness journey. Through sophisticated prompt engineering, rich data context, and a dynamic feedback loop, we deliver a truly personalised and effective coaching experience. Ready to experience truly personalised fitness coaching? Join the waitlist and start your ascent today. Join the waitlist