Most people who do word processing or typesetting use a WYSIWYG word processor or a typesetting system in which they type explicit markup instructions which tell the typesetter how to position text on the page (such as TeX and troff).
Both of these approaches suffer from a few serious problems. The biggest one is longevity of the document: eternal information (the profound things you type) is interspersed with information that will be obsolete (the typesetting information).
Another big problem with this old approach is lack of structure: the markup did not express content, but rather page layout. Let's say you are interested in indexing a bunch of papers written in TeX. It would be rather easy to index all occurances of boldface text, but that's not interesting at all! Instead, it would be really useful to index all function names in an API. With old typesetting approaches you would need artificially intelligent software that could understand the text and say ``aha! this must be the definition of a function in the API''.
So your old world view of writing a document and having the main challenge be how to mark it up to look good on paper is a poor one. Your challenge should be how to mark your document up to emphasize semantic content.