This is awesome! I’ve been working on and off for the last few months on an AI running coach app. Would love to compare notes of effectively prompting Claude or O3 to analyze workouts / build workout plans. In my experience it still requires a lot of hand holding (sudden mileage increases w/w, inconsistent intervals, etc)

Hey - if you're interested in prototyping with real data, we've built a platform for stuff like this -- Fulcra Context (https://www.fulcradynamics.com/). We have an iOS app that collects Apple Health data, and you can then share it with a coach/trainer/doctor/etc.

We have a REST API with python client lib (https://fulcradynamics.github.io/developer-docs/), an MCP server (https://fulcradynamics.github.io/developer-docs/mcp-server/), and did some experiments with ChatGPT actions as well (https://www.fulcradynamics.com/chatgpt).

It doesn't have the simplicity of a plain export to duckdb/sqlite, but it definitely helps solve a lot of the sharing/collaboration problems.

I haven't gone deep on prompting around health data specifically, but I've built some other data agents. Agree that there isn't always an easy path, depending on the domain. It might be worth exploring multiple subagents with different focuses. I.e., have one subagent that is going to focus on writing window function queries in order to find interval variations.

My general belief around data agents is to offload as much as possible to SQL.

i am trying todo same with glue-ing mcp servers together and as a bonus try get my CHO rates and overall KCAL for the week with the proposed schedule and all palnned intake based on mets. Analyzing my workouts is something that works great at the moment if the data in is correct. i mostly use intervals.icu for this.

Would love to chip in on the discussion around this topic.