If I ask 10 of the best developers on the planet to build me something and describe it in great detail, they will all build wildly different things unless I actually point them at the exact thing. And even then, the intent might be identical and the surface might be identical, but everything else in between will be different. Oh, and quality will still vary even in this best base scenario.

That's going to be even worse with LLMs. We would need the exact LLM model and its overall configuration (for open weights), the exact prompts and then the luck of the draw.

We should still encourage open prompts, but it's not a panacea.