In terms of speed and accuracy of intention transfer, normal human conversation proves to be very efficient: exchanged messages carry only sufficient information relative to contextual knowledge assumed to be present at the receiver's end. Furthermore, by receiving layered feedback from the recipient, the speaker is able to verify at an early stage of communication whether his intentions are being accurately perceived. Finally, in divergencies from the expected messages, the listener may ask for clarification at an early stage of message interpretation. For user-system communication to become similarly more efficient, machine interfaces should display both early layered (I)-feedback about partial message interpretations as well as layered expectations (E-feedback) about the message components still to be received. Examples of interfaces are given which already possess these desirable characteristics in part. The "layered-protocol model", proposed by Taylor (1988, Layered protocols for computer-human dialogue. I. Principles, International Journal of Man-Machine Studies, 28, 175-218) as a framework for user-system interface design, details the use of layered I-feedback and related repair messages in user-system communication. In this paper we suggest that the model can be improved by providing it with layered E-feedback, as derived from assumed intentions and layered knowledge of the interaction history.