How does it relate to the posterior as prior question?
As far as I understand the potential is only an alternative way of defining the likelihood. The only way I could see a use here is if you could define an empirical, custom potential from your posterior samples… (you’d need to use an uninformative prior such as a broad enough Uniform distribution).
Say by interpolating from the empirical cumulative distribution.
By the way in the unidimensional case there is the Interpolated distribution that you can use directly as prior. But it does not work in multiple dimensions (Parameterize a multi-dimensional pymc.Interpolated distribution).