LLM “personas” seem somewhat useful for tuning output based on need, I.e. senior SWE doing code review on an established codebase vs. vibe coding proof concept demos, but you’re still right… you’ve only narrowed down the set of stochastic outcomes, not collapsed the output to a true senior SWE