get out of the way
think of ai as a spinning pottery wheel. guide the clay gently.
in an abstract space of problems and solutions, harsh restrictions increase the problem's surface area while reducing the solution's share of the space. sounds sub-optimal.
counterpoint: what if you actually want a super specific solution? details are essential for molding our clay pot. without them, there is just a spinning blob of clay. but the point is to be hyper vigilant in finding and shedding all extraneous restrictions.
The Only Truth re-emerges: balance (stolen from my dad). over-generalization is bad because we risk hallucinations being roped into the huge solution space. over-specificity is bad because we risk excluding the optimal solution from the tiny solution space. the key is in finding the right balance.
example: spend minimal tokens on things like tone, role, expertise, etc. that are tangential to the solution. instead, shape the content of the prompt in a way that shows ai the same clay pot that you have in mind. use jargon. by honing in on the actual content of the problem/solution, it's easier to trust that ai will tap into the right expertise.
further research: Did "Act as a..." only work because the rest of the prompt did not establish the proper ecosystem?
hypothesis: establishing ecosystem with content will yield better results than explicitly restricting to a certain tone/expertise.
method: for math benchmarks, include a random excerpt from a math textbook (or khan academy, etc.) before asking the question at hand.