The insight about prompt format as rhetorical stance realy shifts how we should think about this. Its not just about optimizing for speed or tokens, its about defining the colaboration dynamic you want with the AI. The 88% performance difference between approaches at scale is wild. Most people are probably using JSON thinking its more machine friendly when its actually triggering a compliance mode that bloats the output.
Very interesting research! I might add that prompt format does not only shape the AI’s output, it also shapes the user’s thinking.
Structured prompting helps the user think about the problem and structure their thoughts (and needs)… to select and include the most relevant data (role, context, format, etc.).
The insight about prompt format as rhetorical stance realy shifts how we should think about this. Its not just about optimizing for speed or tokens, its about defining the colaboration dynamic you want with the AI. The 88% performance difference between approaches at scale is wild. Most people are probably using JSON thinking its more machine friendly when its actually triggering a compliance mode that bloats the output.
Very interesting research! I might add that prompt format does not only shape the AI’s output, it also shapes the user’s thinking.
Structured prompting helps the user think about the problem and structure their thoughts (and needs)… to select and include the most relevant data (role, context, format, etc.).
Yeah. I agree. I think structure is even more important on the human side of things. That's enough value in and of itself.
This was fascinating. Thanks for reporting back on that experiment.