I suppose that is true. You can even train the prompt with gradient descent. But in practice, it ends up being fairly different.