We propose a very specifically constrained virtual machine design for goal-directed natural language generation based on a refinement of the technique of data-directed control that we have termed “description-directed control”. Important psycholinguistic properties of generation follow inescapably from the use of this control technique, including: efficient runtimes, bounded lookahead, indelible decisions, incremental production of the text, and inescapable adherence to grammatically. The technique also provides a possible explanation for some well known universal constraints, though this cannot be confirmed without further empirical investigation. In description-directed control the controlling data structure is the surface-level linguistic description of the very text being generated. This constituent structure tree is itself generated depth first by the incremental realization of a hierarchical description of the speaker's communicative goals (neutrally termed a “realization specification”) organized according to the scope and importance of its components. The process of traversing the surface structure gates and constrains the realization process; all realizations are thus subject to the grammatical constraints that accrue to the surface structure at which they occur, as defined by the grammatical annotation of the surface structure tree.