Case study
Product management, applied toAI-built software
A coaching practice's workflow, built into software by directing AI agents
My practice's functional-medicine coaching platform went from requirements to production in about three weeks, built by directing AI coding agents, with a human check on each AI feature.
At a glance
- From requirements to production in about three weeks.
- 300+ commits in its first six months.
- Three AI features, each with a human check.
- It serves the practice's paying clients.
The problem
My functional-medicine coaching practice needed a platform that fits how functional medicine works: root-cause intake, labs read against functional and conventional ranges, supplements and wearable data. Before it, the practice ran on Power BI dashboards and health-coaching data pipelines I had built.
What I found by going to look
The specs came from how the practice works day to day, starting with client calls. They built on health-coaching work that began during COVID.
What I borrowed, and from where
Software product management, applied to AI coding agents. I wrote the specs, ran a ranked roadmap, tested in production and approved every release. AI coding agents (Claude Code) wrote the code under that direction.
What changed
- From requirements to production in about three weeks, on Azure under Microsoft’s BAA.
- 300+ commits in its first six months.
- It serves the practice’s paying clients.
- Features include a root-cause intake with red-flag alerts, labs with functional and conventional ranges, a supplement library with interaction checks, and Oura, Ultrahuman and Withings data.
- It replaced the Power BI dashboards and data pipelines I had built earlier.
Who made it work
- A person checks each AI feature. Three Claude-powered features fit into the practice’s workflow, each with a human check: call briefs that the coach releases, supplement-label readings that staff verify, and estimated diet macros that staff can check.
- Safeguards built in. MFA, role and per-record access, audit logs, notifications that carry no personal health information, and commit hooks that block client data.
- Clear roles. The AI coding agents wrote the code. I wrote the specs, tested, and approved every release.
What carries over today
The same rules work inside any organization. Put AI where a step can be removed or shortened. Keep a person checking the output that matters. Measure the result against how the work ran before. That is the AI half of a Field Review.