How I AIHow to create your own AI performance coach: Optimizing your nutrition, recovery & injury management
CHAPTERS
- 0:00 – 7:07
From lifelong athlete to “I need a system”: injuries spark an AI experiment
Lucas shares his athletic history, repeated injuries, and the motivation to feel “25 in a 40-year-old body.” He explains how the real problem wasn’t a lack of data or experts—but the inability to synthesize everything into clear, daily actions.
- •Competitive sports lifestyle leading to multiple surgeries and chronic issues
- •Aging increases the cost of mistakes in training and recovery
- •Health guidance and metrics are plentiful but fragmented
- •ChatGPT becomes a way to aggregate inputs and drive actionable decisions
- 7:07 – 9:57
Why health optimization breaks down: siloed experts, missing links, and no synthesis layer
Claire and Lucas unpack why even proactive, well-informed people struggle: advice comes from disconnected specialists who don’t share context. Lucas highlights how gaps (e.g., biomechanics, technique changes) can derail recovery despite good medical care.
- •Experts operate in silos (doctor vs PT vs coach vs wearable data)
- •Missing link examples: biomechanics and technique adjustments
- •Need for a unifying strategist to connect recommendations
- •AI positioned as a synthesizer rather than a replacement for clinicians
- 9:57 – 12:09
Inside the WellCoach GPT: the data stack (MRIs, X-rays, wearables, journals, labs)
Lucas screenshares the custom GPT and walks through the files he uploads to give the model deep personal context. The setup spans imaging, structured wearable exports, subjective journal entries, historical labs, and professional nutrition guidance.
- •Imaging: MRIs and X-rays for structural context
- •Whoop CSVs for cycles/readiness/strain and sleep breakdowns
- •Journaling for stress, anxiety, sauna, compression therapy, and daily notes
- •Bloodwork over time to compare trends
- •Nutrition plan and InBody scan to anchor fueling and composition goals
- 12:09 – 13:50
Multimodal + multilingual reality: dumping messy inputs and letting the model unify them
Claire emphasizes the practical breakthrough: heterogeneous formats don’t need perfect cleaning to become useful. Lucas adds that some records are in Portuguese, yet the model can still interpret and integrate them into coaching decisions.
- •Mixed formats: PDFs, images, CSVs, and text plans
- •Multilingual medical records handled without manual translation
- •Lower friction: upload first, organize later (if at all)
- •Imaging adds ‘under-the-hood’ constraints to performance advice
- 13:50 – 15:38
Core instructions: defining the coach’s role, priorities, and success metrics
Lucas explains the top-level system prompt: the GPT acts as a performance strategist for a busy operator balancing training with work. The objective is joint safety, pain-free movement, and sustainable high output—avoiding extreme athlete-mode optimization.
- •Role framing: “performance strategist and health optimization coach”
- •Context: tennis + lifting + recovery while running a company
- •Primary goals: safeguard joints, amplify output, extend peak
- •Action style: interrogate data, flag early warning signs (red/yellow zones)
- •Preference for high-ROI, evidence-based recommendations
- 15:38 – 17:50
Designing for realism: expectations, accessibility, and avoiding ‘biohacker’ rabbit holes
They discuss how strong prompts balance ambition with constraints: practical interventions that fit daily life. Lucas explicitly avoids recommendations like hyperbaric chambers or novelty protocols unless they’re accessible and well-supported.
- •Reasonable targets outperform “be the most elite” prompting
- •Constraints prevent exotic or impractical recommendations
- •Focus on proven, accessible habits and tactics
- •Prompting as a way to keep AI aligned with your lifestyle realities
- 17:50 – 21:48
The operating framework: nutrition, training load management, recovery, and feedback loops
Lucas lays out the pillars that steer the GPT’s decisions: stick to a fueling plan, manage load to prevent overtraining, treat recovery as mandatory, and continuously cross-check advice against the integrated dataset.
- •Nutrition as fuel: stable glucose, low inflammation, muscle retention
- •Training balance: strength, endurance, mobility with joint protection
- •Load gating via readiness/HRV and recovery signals to prevent burnout
- •Recovery stack: sleep first, plus PT/mobility/sauna/cold/massage/mindfulness
- •Cross-validation: recommendations must align with the user’s uploaded data
- 21:48 – 24:55
Hard boundaries and anti-prompts: what the coach must never do
Lucas describes explicit guardrails: don’t push intensity when recovery is low, avoid dubious supplements and novelty interventions, and treat warning signs as reasons to de-risk. The goal is consistency and injury prevention over maximal intensity.
- •No hard training when wearables show red/yellow recovery status
- •Avoid unproven supplements and trendy ‘hacks’
- •No novelty-first recommendations; prioritize measurable ROI
- •Escalate caution when sickness/overtraining signals appear
- 24:55 – 30:30
Real-life example: planning nutrition around an omakase dinner (and other nights out)
Lucas shows a practical workflow: he tells the GPT about an upcoming celebratory meal and gets a day plan that adjusts macros earlier to minimize downside later. Claire highlights how this creates an affordable, on-demand feedback loop for habit reinforcement.
- •Pre-event planning: high protein/low carb earlier to offset indulgence later
- •Using photos of meals for quick feedback and tweaks
- •Adapting plans in real time when social plans change (late night drinks, etc.)
- •Accessibility angle: coaching-like guidance without premium costs
- 30:30 – 32:55
Injury management workflow: elbow recovery, expectations, and tournament planning
Lucas demonstrates using the GPT to track an elbow injury: he inputs diagnosis details, PT prescriptions, pain updates, and media (photos/videos). The GPT helps validate guidance, set realistic timelines, and translate recovery into a stepwise plan toward a competition date.
- •Feeding medical context: diagnosis + PT plan + daily pain notes
- •Using images/videos to localize pain and describe range of motion
- •Validation: AI aligns with doctor/PT guidance rather than contradicting it
- •Expectation management: checkpoints, timelines, and go/no-go decisions for a tournament
- •Reduced anxiety through clearer, more digestible planning outputs
- 32:55 – 37:52
Best practices distilled: multimodal inputs, expert validation, and reformatting medical advice
Claire extracts generalizable tactics: use visuals, ask AI to sanity-check expert recommendations, and have it translate instructions into the format you’ll actually follow. This turns one-time clinical guidance into a practical daily execution plan.
- •Photos/videos + annotations can be high-signal inputs
- •AI as an always-available second pass to clarify expert guidance
- •Reformatting: convert clinical instructions into day-by-day plans
- •Improving adherence by making plans easier to understand and execute
- 37:52 – 43:52
The near future: ambient health data capture and AI-to-AI doctor/patient collaboration
Lucas outlines a vision where personal AI coaches are ubiquitous within five years, complementing—not replacing—doctors. He predicts seamless data capture via sensors and a future where institutional medical expertise becomes productized into trusted, personalized models.
- •Personal AI coaches to support between-visit decision-making
- •Doctor AI + patient AI exchanging context before appointments
- •Shift from manual logging to passive, ambient tracking
- •Potential sensors: glucose/hormones, smart fabrics, toilet/microbiome signals
- •Institutions may license expertise as trained models for personalized guidance
- 43:52 – 49:17
Other AI workflows: synthetic clients and an ‘AI co-founder’ for distributed teams
Lucas describes using AI to approximate busy stakeholders: modeling how a client-doctor thinks (using only public information) and creating an AI brainstorming partner to reduce coordination overhead in a remote company. Claire connects this to synthetic bosses/clients as a productivity multiplier.
- •Synthetic client: approximate expert feedback when the real person is unavailable
- •Guardrail: no proprietary client data—public sources only
- •AI co-founder: brainstorming partner for thorny problems without scheduling meetings
- •Distributed work increases the value of on-demand ‘thinking partners’
- 49:17 – 51:36
Reliability, guardrails, and closing: treating AI as evolving tech and staying grounded
They close on how Lucas handles occasional bad outputs: he relies on guardrails, limits external info, and views mistakes as part of model evolution. The episode ends with where to find Lucas (via his company) and final sign-off from Claire.
- •Models are improving: fewer hallucinations over time
- •Guardrails and constrained context reduce off-track advice
- •Healthy mindset: treat AI like a developing tool, not an oracle
- •Contact: Lucas via Cactus.is; playful close about tennis challenges
