1) When I read #uri "http://google.com" my app code sees (java.net.URI. "http://google.com"), not Tag "URI" "http://google.com" or whatever. clojure.core/map does not see tagged values, it does not know that the values were ever read from edn.

2) Extension happens in userland library code, you don't need to go modify the hardcoded pattern match in core. (Talking about reifying actual instances that we can code to, not reader tags)

3) Data is just information. Information isn't coupled to code, it's abstract values, totally separate from the concrete implementation. As RH says: "code is data. But data is not code until you define a language around it." Typeclasses are about code.

4) EDN values are cross platform and a platform's EDN reader can reify the value into an idiomatic type for that platform. E.g. a haskell edn reader could reify #error "foo" into Left String; a Java reader a Throwable to be re-thrown later.

5) The whole prism diversion is sophomoric. Once you've read the EDN into concrete values of whatever platform type, you can use whatever platform abstractions you like to manipulate them. Clojure has lenses too: http://funcool.github.io/lentes/latest/#composition

You can watch the EDN talk or read the transcript if you'd like to learn more. This topic is very deep but this thread is not doing it justice.