Ergatta - Custom GPT & MCP (Model Context Protocols)

May 21

The Opportunity

With the advancements in AI, I was looking for opportunities to optimize our content creation pipeline for maximum impact with minimal overhead.

This was targeting the workout content team as an internal tooling workstream.

Hands-On Product Manager / Prompt Engineer:

Iteratively developed prompts teaching Claude how our workout content is structured, provided constraints and exercise science guidelines for how to design Ergatta interval workouts.
Shared JSON payloads from existing workout content with Claude so that it understood the output I was expecting from it.
Feed Claude our existing library of interval workouts, to know what exists so that it would ensure it’s workouts are unique, but also to learn from how we design workout content.
Iterated through testing, debugging, and optimizing my prompts to include validations and adding check points for human approval.
Packaged this up into simple process that others on the team could use to prove out the use case and learn from it.

The process automated the creation of workout content for 7 different game experiences on the Ergatta platform.
Reduced workout development time from roughly 10 minutes to about 1 minute per workout.
Increased workout creation from 24 workouts to over 100 workouts per month.

This first phase was a proof of concept and learning opportunity.
AI develops code from scratch to execute the prompt and needs to iteration to get that code working as expected (correcting mistakes it made).
Validations provided in the prompt do not matter unless AI had built code to perform those validation checks. This required edge case testing to prove each validation was in place, otherwise correcting the AI model to add or modify the validations.
I exported data from out database as a CSV file in order to keep the process silo’ed, however the success here is encouraging us to looking into expanding it’s capabilities further with Model Context Protocols where we can give it more autonomy to execute the process with a little human oversight.
Future plans are to take our Workout Recommendations engine and use that as the input for curating content tailored to target persona to boost their engagment.

Christopher Chang