Back to the "Quotes and punctuation" problem: So the main problem in TeX which makes pattern matching/lookahead/parsing very difficult in TeX is that the parsing routine cannot know into what kind of stuff future tokens expand without literally expanding the whole document. Would a two-tier expansion mechanism (to be implemented as an extension to TeX-the program) help? I am thinking along these lines: Both macro names and each of their arguments would get one of the attributes TRANSPARENT (T) or OBLIQUE (O). One could then have a first-tier expansion (T-expansion, say) in which only those macros and arguments which have the T feature are expanded. The result could then be processed with traditional methods (regular expression matching, tansformation patterns etc.) before final expansion of the O-macros takes place. Below I indicate the expansion categories for a few standard LaTeX macros in a sort of Pascal like fashion, i.e., the first category refers to the macro itself, the others to the arguments: TRANSPARENT \ref{TRANSPARENT} OBLIQUE \section[OBLIQUE]{TRANSPARENT} OBLIQUE \label{OBLIQUE} OBLIQUE \newcounter{OBLIQUE}[OBLIQUE] OBLIQUE \flushbottom OBLIQUE \begin{tabular}[OBLIQUE] TRANSPARENT \end{tabular} OBLIQUE \makebox(OBLIQUE,OBLIQUE)[OBLIQUE]{TRANSPARENT} TRANSPARENT \cite{TRANSPARENT} OBLIQUE \bibitem[OBLIQUE,OBLIQUE] OBLIQUE \emph{TRANSPARENT} OBLIQUE \itshape TRANSPARENT \newcommand{OBLIQUE}[OBLIQUE][OBLIQUE]{OBLIQUE} Assume, e.g., that the \ref command is defined to typset boldface, and a command \last has been defined which expands into "last". Let's T-expand the line Theorem \ref{fermat} was his \emph{\last}. Since \ref is TRANSPARENT, it expands, while \emph is OBLIQUE, but has a TRANSPARENT argument, which expands. So the result may look like Theorem \textbf{1.3} was his \emph{last}. A pattern matcher could now e.g. ignore all OBLIQUE tokens in this text, and thus match on "his last". How will this help with the quote problem? The text \newTcommand{\story}[T]{There was a man, who said, <I'll tell you a story>, and he began: <#1>} \story{\story{\story{\ldots}}}} will T-expand into There was a man, who said, <I'll tell you a story>, and he began: <There was a man, who said, <I'll tell you a story>, and he began: <There was a man, who said, <I'll tell you a story>, and he began: <\ldots>>> I assume that < and > are active characters of category OBLIQUE, so they do not expand at this stage. When full expansion takes place, the > can do a lookahead and, e.g., detect the following > from which it can determine optimal spacing (keeping track of the nesting level should be easy, I guess). The problem with braces (e.g. when a font change takes place) does seem linked to the particular way that \futurelet works, but I don't see how this poses a fundamental problem with a lookahead-type strategy. Two comments regarding other points that came up in the discussion: While I think that this quote problem might be useful to illustrate general parsing problems, I don't think it is reasonable to expect any author to write \quote{Quoted text}, or to make super smart decisions about the placement of punctuation. Any author who cares will do it right, and those who don't care won't want to write \quote, anyway. Related to this: Yes, I do believe that the primary goal of LaTeX should be to provide a human readable direct input typesetting language. If the abstract structure of the document can be completely specified, the better, but it seems that these two goals may be ultimately incompatible. I really know SGML/XML (Sebastian, what do you hope to get out of the LaTeX3 project that you think you cannot get out of SGML?), but I seem to understand that SGML is well suited to provide complete logical markup. So why re-do something that already exists and supposedly works well? Moreover, especially when typing Mathematics, complete logical mark-up is way beyond what most authors need in practice. If I need a tool which is optimized w.r.t. typsetting, I use LaTeX. I don't care whether the mark-up is sloppy in certain places---it is not important to me whether I can re-use the input in more structured applications. An example: for humans it makes perfect sense to use \ldots in a formula, but I could not even expect a symbolic system like Mathematica to understand what I mean. If my requirements go beyond publishing, my primary tools are different (although it would help if they could export to LaTeX when it comes to publishing), but then I don't mind the extra effort required. Marcel