One problem is that way of thinking is distinctly ASCII-centric -- what happens when you have accents, umlats, Chinese, and the rest of the stuff that comes out

One problem is that way of thinking is distinctly ASCII-centric -- what happens when you have accents, umlats, Chinese, and the rest of the stuff that comes out of Unicode? I suspect the encoding you propose would actually increase the size of the average plaintext, simply because people rarely will type "Hamlet" and then later type "HaMlEt" in the same document. I think a far more common case would be "Hamlet" (beginning of sentence) and "hamlet" (inline), where you can reuse the substring "amlet". (Either that, or they're consistent with the usage of all-caps, like MICROSOFT(R).) Besides, if you know you're doing English, you can save 12.5% simply by using 7-bit ASCII instead of of 8-bit ASCII.