“Learning” by prompting, calculating the loss against evals, and updating the prompt