I do think there is some actual value in telling an LLM "you are an expert code reviewer". You really do tend to get better results in the output

When you think about what an LLM is, it makes more sense. It causes a strong activation for neorons related to "code review", and so the model's output sounds more like a code review.