Early cognitive models of spelling assumed that orthographic word representations are linear, ordered sequences of abstract letter identities (graphemes), activated only by word meaning information, and in some cases proposed that activating phonological information is a necessary stage of the spelling process. Over the past 20 years, studies on dysgraphia have shown that orthographic representations are autonomous from phonological representations and, just like the latter, are directly activated from semantics. The selection of an orthographic form for output relies on the convergence of activation from lexical-semantic information and from sublexical phoneme-grapheme conversion procedures. In addition, it is increasingly clear that orthographic representations are multidimensional objects that separately represent the graphosyllabic structure (or perhaps the nucleus/non-nucleus positions) of the target, and the identity, the CV status, and the quantity (doubling) of each grapheme. In spelling, the structure of orthographic knowledge and the mechanisms involved in processing serial order interact in complex ways and constrain performance accuracy. Further research is needed to clarify some critical issues: We need to specify in greater detail the mechanisms involved in the interaction between meaning and sublexical information; we must consider the possibility that orthographic representations have texture, in addition to structure; we must provide explicit hypotheses on the mechanisms that process orthographic knowledge; and we must gain a better understanding of the interaction between structure and serial order.