If the human is involved, "where what the human is saying to guide the model", then the prompts are even more important to include. Is this a case of "solve an open problem in theoretical physics, make no mistakes", or a case where a human had to interact frequently with the model?