Q4C12.XMLDesc and its submodules provides a way to describe the structure of an XML document and how to map it to and from a Haskell value. A RELAX NG schema can then be generated from the description. Custom datatype libraries (in the RELAX NG sense of the word) are supported.
There is one minor footgun: if your schema is recursive, then you must use one of the non-terminal functions to give a part a name and to break the cycle. Otherwise, it works fine as a parser and a printer, but a RELAX NG schema cannot be infinite.
Q4C12.XMLDesc uses an Applicative-like approach, except instead of having fmap :: (a -> b) -> f a -> f b, it has rfmap :: Iso' a b -> f a -> f b, where Iso' is a (full) isomorphism: it can go in either direction losslessly. (Note that 'lossless' may be context-dependent: for example, ignoring space at the start and end of a field is not an isomorphism over all strings, but it may be suitable for a specific application.) This is similar to the idea of using partial isomorphisms (outlined in, e.g., this blog post and originally described in this 2010 paper by Tillman Rendel and Klaus Ostermann), but without the partiality.
In addition, the van Laarhoven representation of isomorphisms is used, allowing interopability with the lens library, and reducing the need for Template Haskell.
Using the Desc typeclass for an abstract structure definition and then having separate print, parse and generate-the-schema instances is outlined by Joachim Breitner in a blog post; it also owes an obvious debt to the inimitable Oleg and his work on tagless final interpreters.