Another way of thinking about this may be as a compressed field, where the "prompt" extracts knowledge along a given vector, along the lines of stored inter-token relations, which provide for the internal vectorization, thus reducing redundancy and enabling the compression in the first place.