Sure, you can do that. The parent's point is that if you want this mapping to obey the rules that an actual definition in (say) first-order logic must obey, you run into trouble. In order to talk about definability without running into paradoxes, you need to do it "outside" your actual theory. And then statements about cardinalities - for example "There's more real numbers than there are definitions." - don't mean exactly what you'd intuitively expect. See the result about ZFC having countable models (as seen from the "outside") despite being able to prove uncountable sets exist (as seen from the "inside").