I skipped to the conclusion to get the overall jist. I think these sorts of studies are important going forward; the prompts efficiency in making for clear, guided instructions the ML can then efficiently handle is exponentially important as context grows. If you are asking the right questions the right way, you can get more back. But if you are causing the LLM to struggle routing its chain of thought it will only get so far along until you can no longer extract anything useful from it (and may well be getting gaslit in the process).
Obvious to some, but it's definitely overlooked by many.