As regards how often I do this: I go through periods where I do it a lot (sometimes several times in a week), and then months where I don't do it at all. I do it thematically (i.e., with closely related papers), so I've never tried doing something like what you suggest with the arXiv's recent papers. They're usually not all that closely connected.
In addition to Arxiv and preprints I can find online, there's Google Scholar and Amazon Previews (I'm still missing many journal articles, especially in engineering, due to a lack of university access, luckily they're often compiled in journals on Amazon). By flipping through Amazon's book previews using the search feature, I can read an arbitrary number of pages in any given book, and the world's library is at my lap. I can then 'photocopy' the relevant/interesting sections using ctrl-shift-command-4 on my Mac, and paste them into my Evernote. In this way I can locate and collate a large number of papers and texts, and organize them along the way without even dipping into LaTeX. After that, I can past those copies into Mathematica, which has a very workable equation typesetter, with the additional advantage of the equations being computable.
Lately I make a lot of use of Evernote, however, which I can pretty much paste anything into, and I can 'photocopy' any part of a text on the computer by using ctrl-shift-command-4 on my Mac.
http://bytepawn.com/readings-in-distributed-systems/
But, more importantly, I'm curious about the 15/3-routine. How often do you do this? Recently I subscribed to the arxiv RSS (astro-ph.co, gr-qc), there's lots of papers uploaded daily, but most don't look very interesting, so 15 interesting ones would be ~1 weeks worth for me.