As the name prior says, it incorporates your (subjective) prior knowledge about the parameters under consideration. It must not depend on the data! It is completely up to you to set it. Bad, but common practice is to use the priors of seminal papers like Smets/Wouters (2007) or Justiniano/Primiceri/Tambalotti.
Please also see [DSGE Estimation - Imposing prior mean and std) and [Choosing an appropriate standard deviation for prior)