> Couple this with the fact that models respond better/worse to certain prompts depending on the stylistic composition of the prompt itself.

Do we really know this, or is it just gut feeling? Did somebody really proved this statistically with a great certainty?