ThePrimeagen experiments with Jev, building a TypeScript agent loop to play Balatro by transforming raw game state into reduced prompt payloads and structuring sequential action decisions.
Original by The PrimeTimeAgent workflowsIntermediate148 min 40 secPublished
The coding session reduces a large game-state payload before sending it to Jev, making the relevant hand and scoring data explicit.
Each request needs the necessary state again; the stream does not treat Jev as a conversation with automatic memory.
The prototype divides play into smaller state-machine or behavior-tree choices rather than asking for an open-ended multi-step plan.
Worth knowing
Decision latency and model reasoning quality heavily depend on prompt reduction and external game state extraction logic rather than fully autonomous general gameplay capabilities.