If you think in terms of the operations that perform the tree transformations, of course you will have difficulty because, you're using a different tree: the syntax tree representing the code of those transformations! Not the syntax tree of the source and destination structure. So you are "one removed" from the real problem. Moreover, since both domains are trees, it's confusing.

It sounds like you could have benefited from a way to express pattern matching using JSON syntax, for destructuring a JSON object, and to represent the synthesis of new JSON using a "JSON quasiquote".

E.g. Lisp transformation with classic destructuring-bind and backquote:

    (destructuring-bind (a (b (c &rest d) e) f (g h)) some-obj
      `((,a ,b) ,c (,@d 1 2 3 ,e) (,f ,g, h)))

There are pattern matching libraries nowadays that do a lot more than destructuring-bind, but it illustrates the basic point.

The destructuring-bind macro writes the code to pull apart some-obj according to the tree picture with embedded variables. The backquote syntax generates the code whose evaluation synthesizes the new tree object according to a template, with the values of expressions indicated by , and ,@ substituted and spliced into the template.