A comprehensive investigation is carried out to establish best practice guidelines for the modelling of statistically steady-state non-neutral urban boundary layers (UBL) using large-eddy simulation (LES). These steady-state simulations enable targeted studies under realistic non-neutral conditions without the complications associated with the inherently transient nature of the UBL. An extensive set of simulations of convective and stable conditions is carried out to determine which simplifications, volumetric forcings, and boundary conditions can be applied to replicate the mean and turbulent (variance and covariance) statistics of this intrinsically transient problem most faithfully. In addition, a new method is introduced in which a transient simulation can be ‘frozen’ into a steady state. It is found that non-neutral simulations have different requirements to their neutral counterparts. In convective conditions, capping the boundary-layer height h with the top of the modelled domain to h/5 and h/10 (which is common practice in neutral simulations) reduces the turbulent kinetic energy by as much as 61% and 44%, respectively. Consistent with the literature, we find that domain heights lz≥5|L| are necessary to reproduce the convective-boundary-layer dynamics, where L is the Obukhov length. In stably stratified situations, the use of a uniform momentum forcing systematically underestimates the mechanical generation of turbulence over the urban canopy layer, and therefore leads to misrepresentations of both the inner- and outer-layer dynamics. The new ‘frozen-transient’ method that is able to maintain a prescribed flow state (including entrainment at the boundary-layer top) is shown to work well in both stable and convective conditions. Guidelines are provided for future studies of the capped and uncapped convective and stable UBL.