By "manual" I mean various scripts, on MacOS. "bib" for example is an Alfred keyword that runs a script scraping the front browser page for bibtex data, and initializing a new folder based on this data. Other scripts write search output files that "Find Any File" thinks it saved; FAF offers a nice Finder-like interface for selectively opening search results.
I've avoided any canned solution. I don't want vendor lock, and I'd rather be crippled by my own lack of imagination than someone else's.
By far the most useful metadata are the dates of access. One wanders continuously through an idea space no software can adequately categorize, yet adjacent items in time are likely related.
The holy grail I haven't coded yet is to generate web pages with URI schemes that open each PDF in a viewer. GoodReader is a nice viewer on iOS with such an interface; stunningly its own browser can't understand these links, but DropBox can. To simplify versions for MacOS and iOS, one can write a custom URI handler on MacOS that recognizes the GoodReader link, and redirects it on a Mac.
The goal here is to meaningfully browse closer to the speed of light than clumsy current standards. Just as lock-picking is both skill and cycles, research is both skill and cycles: One wants to have the right three ideas in close enough proximity for our feeble brains to notice the connection. We have to be in our happy place when we're flailing for hours, days, years; it would nevertheless be nice to accelerate this process.
I've been in constant friction with the MathSciNet team: I believe that it should present itself as the premier playground for machine learning mind mapping experiments, in addition to maintaining its stodgy hand curation. There is some controversy as to how math is organized. Many great mathematicians wander freely; others never read a paper outside their field after the age of 23. Departmental hiring meetings degenerate to "it's the number theory group's turn", perpetuating a dated view of how math is organized. Independent views of the MathSciNet database could give us new understandings of our field.