[kbmr]

>Reviewing LLM output requires constant context-switching between "what does this code do" and "is this what I actually wanted."

Best way I've seen it framed