> who prompts the AI

LLMs are a box where the input has to be generated by someone/something, but also the output has to be verified somehow (because, like humans, it isn't always correct). So you either need a human at "both ends", or some very clever AI filling those roles.

But I think the human doing those things probably needs slightly different skills and experience than the average legacy developer.

Rules engines were designed for just such a thing. Validating input/output. You don’t need a human to prompt AI, you need a pipeline.

While a single LLM won’t replace you. A well designed system of flows for software engineering using LLMs will.