How to create your own AI performance coach: Optimizing your unique nutrition, recovery, and injury management needs | Lucas Werthein (Cactus)

Lucas walks through the technical setup of his custom GPT, including the diverse data sources he integrated: X-rays and MRIs of injuries, Whoop wearable data (sleep, strain, recovery), blood test results in multiple languages, nutrition plans, and InBody scans. He demonstrates how ChatGPT can process unstructured data across different formats and languages without manual cleaning or structuring.

• •Uploaded medical imaging (knee MRIs pre/post-surgery, X-rays), blood tests, nutrition plans, and wearable data as raw files to custom GPT
• •System processes data in multiple formats (CSV, PDF, images) and languages (English/Portuguese) without preprocessing
• •Integration of both structural health data (imaging) and performance metrics (workouts, sleep, strain) provides comprehensive context
• •InBody scans track body composition changes (fat/muscle distribution) to inform training recommendations
• •All data aggregation happens in one place, solving the siloed information problem common in healthcare

Detailed walkthrough of the GPT's instruction set, including role definition (performance strategist for high-performance operator), optimization priorities (joint protection, energy output, peak extension), and hard boundaries. Lucas emphasizes evidence-based recommendations over trendy biohacks, with specific guardrails against overtraining and unproven supplements.

• •Role definition: Act as performance strategist balancing competitive tennis, weightlifting, and running a company - not training a professional athlete
• •Primary objectives: Safeguard joints, amplify output, extend peak performance, feel '25 in a 40-year-old body'
• •Optimization priorities: Stick to nutrition plan unless data-driven reason to adapt, balance strength/endurance/mobility, prioritize recovery as non-optional
• •Hard boundaries: No pushing past volume when metrics show under-recovery, no unproven supplements, no novelty biohacks, flag red/yellow zones immediately
• •Response format: Clear, tactical, no fluff, connect dots across data sources, prioritize what matters this week over vague long-term theory

Lucas demonstrates practical daily use cases, including preparing for an omakase dinner by adjusting meals throughout the day to balance indulgence. Shows how the AI provides specific meal recommendations (protein-focused, minimal carbs) and validates food choices through photo analysis, offering micro-adjustments like adding pepper for inflammation.

• •Proactive planning: Asked how to manage day before high-carb/sake dinner, received structured meal plan (protein-only breakfast/lunch)
• •Visual feedback loop: Took photo of breakfast, AI validated choices and suggested small optimizations (add black pepper for inflammation)
• •Adaptive replanning: When plans changed (party at 3AM instead of dinner), AI adjusted recovery recommendations on the fly
• •Accessibility insight: AI can recommend optimal choices even at limited options like Chipotle, making healthy eating more practical
• •On-demand coaching eliminates need to schedule nutritionist calls for simple daily decisions

Lucas shares how he used the GPT to manage an elbow injury by uploading doctor's diagnosis, PT prescriptions, and daily pain updates via photos and videos. The AI validated medical advice, managed expectations with visual recovery timelines, and helped him plan return to competitive tennis by September 18th tournament.

• •Uploaded doctor diagnosis, PT prescription, and tracked daily pain through photos/videos with circled pain points
• •AI validated professional medical advice while presenting it in more digestible, contextualized format
• •Created visual recovery timeline with decision checkpoints for tournament readiness assessment
• •Reduced anxiety by providing clear expectations vs. daily 'is the pain gone?' uncertainty
• •Demonstrates using video/image inputs with annotations to communicate physical symptoms effectively to AI

Discussion of underutilized AI capabilities for health applications, including circling areas of concern in photos/videos to diagnose issues. Examples include home maintenance problems and body pain localization. Highlights how AI can validate expert opinions, re-explain medical advice in different formats, and provide 24/7 accessible second opinions.

• •Taking photos/videos and circling specific areas is highly effective for communicating symptoms or problems to AI
• •Use case extends beyond health: home repairs, structural issues, any visual problem diagnosis
• •AI serves as on-demand validator of expert opinions when professionals aren't available 24/7
• •Can re-format expert advice into personalized structures (e.g., day-by-day plan until specific date)
• •Helps translate medical terminology and mechanics into understandable language for patients

Lucas envisions a future where everyone has a personal AI health coach that synthesizes passive data from microsensors, smart fabrics, and ambient monitoring. Healthcare institutions will sell trained AI models based on decades of patient data, and patient AIs will communicate with doctor AIs to enable more informed, efficient medical visits.

• •Within 5 years: Personal AI health coaches become standard, helping people show up to doctors more informed and make better daily micro-decisions
• •Seamless data capture: Microsensors in bloodstream, smart fabrics, toilet sensors measuring microbiome/hydration - all passive and invisible
• •Healthcare institutions (Mayo Clinic, etc.) will sell AI models trained on decades of patient data combined with personal health information
• •AI-to-AI communication: Patient's AI will talk to doctor's AI before appointments, providing full context for more productive visits
• •Doctors shift from data entry to storytelling and long-term strategy, freed from manual tracking burden

Lucas demonstrates two business GPTs: a 'synthetic client' trained on publicly available information (articles, podcasts) to approximate client thinking when unavailable, and an 'AI cofounder' for brainstorming thorny problems in distributed work environments. Both serve as on-demand thought partners to accelerate work and reduce dependency on busy experts.

• •Synthetic client GPT: Trained on public information (articles, podcasts) to get work to 80-90% before client review, saving expert time
• •Important caveat: Only uses publicly available information, no proprietary client data to avoid training model on sensitive information
• •AI cofounder: Helps brainstorm problems in distributed teams when you can't just tap colleague's shoulder for quick discussion
• •Serves as 'business therapy' - voice to think through problems without starting from blank slate
• •Pro tip: Bosses/clients should create GPTs of themselves to share with teams for passive first-line feedback

Discussion of how this AI coaching framework applies beyond elite athletic performance to everyday wellness goals, caregiving, youth athletics, and aging well. The same structure can help people with basic goals like eating less processed food, sleeping better, and moving more through personalized nudges and contextualization.

• •Framework applies to diverse use cases: caregiving for elderly relatives, youth sports development, general wellness, aging optimization
• •Most people don't need six-packs or match prep - they need help with basics: less processed food, better sleep, more movement
• •AI coach can meet people where they are with appropriate nudges and contextualization for their specific goals
• •Caregiving example: Synthesizing multiple specialists, data points, pictures, recordings for elderly care management
• •Youth athletics: Same framework can help 8-year-old improve at basketball with age-appropriate training and feedback