Stage 1: Draft generator takes user input and creates high-level workout structure with segments Stage 2: Specialist processors (warm-up expert, interval specialist, etc.) convert each segment into precise power targets and timings
Key insights: - LLMs excel at generating structured JSON when you use schemas - Breaking complex tasks into smaller, focused LLM calls works better than monolithic prompts - Each specialist has isolated context, forcing self-contained outputs
The result: "4x4min threshold intervals" becomes exact power zones and durations that sync directly to Wahoo element bike computers.
Anyone else finding structured outputs surprisingly reliable for complex data generation?