about

I had a “skill” (a saved prompt) in dia that recommended recently released albums to me. After I stopped using dia, I would still occasionally open it just to run that prompt.

The first version of this project gave the prompt a new home by putting the whole process inside a small app: running the prompt on a schedule, gathering album data, and displaying the results. It worked, but it also meant that changing how the agent operated meant changing the app.

This version deliberately separates those jobs. The app is a simple home for finished recommendations, while an agent runs elsewhere and publishes completed sets to it. That leaves me free to change the prompt, model, tools, or cadence—and to run, inspect, or intervene in the process—without building agent infrastructure into the app itself.

The agent can research albums across the wider internet, including reading reviews, but the app needs an authoritative ID for every album it stores. Its API therefore has one endpoint for finding those IDs and another for submitting the finished suggestions. The endpoints and instructions for using them are documented in the app’s llms.txt.