The prompts don't matter if the training data isn't up to par - more importantly I believe the nature of the training is such that the weight activations for the various prompts are unlikely to be independent.

In other words if 20% of your training data is scientific literature, even with appropriate disambiguating prompts the output will still be heavily influenced by the other 80% of your training data.

When you use GPT-3 to generate outputs, you're actually sampling from a learned subset of a super complex, super high dimensional space - and without human knowledge all the neural networks are doing is translating priors (input prompt) into points in the learned space. And the learned space is some complex topology of points between which the net interpolates - it's extremely difficult with current tech to control the shape of this learned space and that shape is influenced by all training data under a scheme like GPT-*.