this appears to be the more complete solution, at least in the case I tested with gpt-3.5-turbo-16k: API Prompt for gpt-3.5-turbo-16k - #12 by SomebodySysop
Until I started sending the documents in XML format, gpt-3.5-turbo-16k would pretty consistently not recognize them if a) there were lots or b) the text was fairly complex. No such problems with gpt-4 or gpt-4-turbo.