Building a Personal AI Wellness Coach to ‘Feel 25 in a 40-Year-Old Body’

By: Lucas Werthein Host: Claire Vo Source: How I AI (ChatPRD) Type: podcast

Summary

How I AI episode with Lucas Werthein, Head of Technology at Cactus (clients include Apple, Coca-Cola, MTV, Beyoncé). Three workflows applying the same AI synthesis pattern across personal health and professional work: (1) Personal AI wellness coach — uploaded diverse, unstructured health data in mixed formats and languages: knee X-rays and MRI scans (pre- and post-surgery images), Whoop strap CSV exports (physiological cycles, daily journal entries, workout logs, strain scores, sleep data), blood exam PDFs spanning multiple years, dietician nutrition plan, InBody body composition scan; configured custom GPT as “performance strategist and health optimization coach” for a “high-performance operator” balancing competitive tennis, weightlifting, recovery, and running a company; defined four optimization pillars (nutrition aligned to existing plan, training/load management respecting Whoop readiness scores, recovery as non-negotiable, all recommendations cross-validated against provided data) with hard boundaries (never push through pain when under-recovered, no unproven supplements, no novelties, act on red flags like low HRV or poor sleep); uses daily for meal planning around social events, managing injury recovery with photo/video updates of exact pain locations, and tournament timeline planning with recovery checkpoints; (2) “Synthetic clients” — create dedicated GPTs for key clients using only publicly available information (articles, podcasts, presentations); team queries synthetic client to iterate work to 80-90% completion before actual client meeting; strictly no proprietary or confidential data; (3) “AI co-founder” — GPT loaded with co-founder’s thinking style and specific business challenges; serves as on-demand brainstorming partner in remote work environment; fills the gap of spontaneous in-person collaboration.

Key Ideas Extracted

Multi-format health data synthesis: Upload X-rays, MRIs, CSVs, PDFs, and images in multiple languages without cleaning or standardization — modern AI models handle unstructured, multi-format data surprisingly well
Four-pillar optimization framework with hard boundaries: Define clear domains (nutrition, training, recovery, feedback loops) AND explicit constraints (never push through pain, no unproven supplements) — guardrails prevent the AI from giving generic internet health advice
Whoop readiness scores as training governor: Configure the AI to respect wearable recovery data before recommending training intensity — prevents overtraining by tying advice to objective physiological signals
Photo/video feedback loops for injury management: Upload photos pointing to exact pain location and videos of movement — supplements text descriptions with visual context the AI can interpret for more precise recovery guidance
Synthetic clients from public data only: Build client persona GPTs from publicly available articles, talks, and podcasts (never confidential info) — gets work to 80-90% before real client meeting, maximizing expensive face time
AI co-founder for remote brainstorming: Load a GPT with a colleague’s thinking style and current business context — simulates the spontaneous “tap on the shoulder” conversations lost in distributed work
AI as synthesis engine, not content creator: The common thread is combining diverse, scattered information into unified, actionable insights — MRIs + sleep data + nutrition plans, or client interviews + articles + presentations
Daily micro-decisions compound: Using the coach for small daily choices (breakfast before a dinner party, whether to train after poor sleep) creates sustained impact over time — the value is in consistent application, not dramatic interventions

Notes

Published Nov 24, 2025 on How I AI (ChatPRD). ~9 min read. (Note: filename says 2025-10-27 but actual publication date is Nov 24, 2025)
Sponsors: WorkOS, Google Gemini
Lucas Werthein background: Head of Technology at Cactus; competitive tennis player; clients include Apple, Coca-Cola, MTV, Beyoncé
Tools: Custom ChatGPT (wellness coach, synthetic clients, AI co-founder), Whoop (wearable data), InBody (body composition)
Data formats used: JPEG/image files (X-rays, MRIs), CSV (Whoop exports), PDF (blood tests, nutrition plan), photos, videos — in English and Portuguese
Key persona instruction: “performance strategist and health optimization coach” for a “high-performance operator”
Three companion workflow guides published Jan 8, 2026
Cross-references: Custom GPT design, health data synthesis, synthetic personas, AI coaching, remote work tools

Raw Content

Re-scraped from ChatPRD 2026-02-16. Full article content captured in Summary and Key Ideas above.