It's amusing how "JSON output" as a line-protocol between LLMs is a hot topic in some circles. Some have rightly pointed out that this is merely a selection of grammars. But more broadly it is a selection of languages. When teaching languages to children ( meat or machine ), one must first teach the child to build internal models, before worrying about external grammars. What does that mean?
Well, for any sensory dataset, there is a "cognition" of its informative value ... ( "difference from null" is the definition of information ) ... this is discussed to a great extent in the study of gestalt psychology, which fortunately for formalists, is very easy to model mathematically.
Now here comes some differentiation. Of the children who can recognise a sensory pattern, NOT everyone can imagine it internally, due to limitations of their conscious sensory imagination : and those who cannot, must use external modelling, via gestures, or pen and paper, to visualise what they want to think about.
However, once a child demonstrates a consistent ability to recognise and manipulate a sensory model, ONLY THEN is there a question of how the child should verbally express it.
This is a fundamental matter in semiotics, and language in general.