Thanks for this comment -- I'll put the second strategy in effect (the first one perhaps I'll leave for later as I currently work in an open plan office).

I wonder if this "comment driven" strategy has potential for applying NLP+code generation. I'm sure this is a common enough thought but I'm not aware of the research, does anyone here have pointers?

What I mean is: use NLP to extract intent (where feasible), and then offer possible implementations of the intent.