For any of the big things I've worked on, it isn't one "prompt", it's a goal that keeps getting updated and a bunch of back-and-forth steering of the model and sub agents, where what the human is saying to guide the model is heavily context dependent on what the model is currently doing or saying. Much of that guidance will ultimately be useless guidance on dead-end explorations. Also, the model will likely have written huge programs to solve things symbolically or simulate things numerically, and both the programs that were written, and the output of those programs are critical to understanding how the model came up with the things that led to a solution.
Now, I'd be all for publishing the full transcript, but the transcript is likely to be massive. Gigabytes of text, and reproducing the local computation (often sympy or numpy) may be hundreds or thousands of CPU-hours.
I wonder how this works out in other parts of physics codes. If you need a Monte Carlo integrator, is it common to guarantee a seed for reproducibility? How do these codes handle distributed systems for rng? Is it common to use a fully reproducible rng and GPU kernel for scientific compute?
But it isn’t “just a prompt”, and sharing it doesn’t make the results reproducible (since you don’t get control over the model’s “seed” from run to run).