A method is presented for controlling the field-directed self-assembly of colloidal magnetic core-shell nanoparticles into three-dimensional (3D) crystalline superstructures with nanoscale feature resolution. This level of resolution is obtained using submicron soft-magnetic template elements to guide the assembly in the presence of a uniform bias field. The use of a bias field combined with template-induced gradient fields is a critical feature of this process as it provides highly localized regions of attractive and repulsive magnetic force that enable nanoscale control of particle placement during assembly. We demonstrate proof-of-concept using a computational model that predicts the dynamics of individual particles during assembly as well as the final assembled structure. Our predictions are consistent with reported experimental observations and demonstrate for the first time that 3D crystalline superstructures can be assembled within milliseconds. Uniform hexagonal close packed (hcp), face centered cubic (fcc), and mixed phase hcp-fcc structures can be obtained depending on the template geometry. We further show that the structure and resolution of the assembled particles as well as the rate of assembly can be controlled via careful selection of key parameters including the template geometry, particle constituents, core-shell dimensions, and the particle volume fraction. The proposed assembly method is very versatile as it broadly applies to arbitrary template geometries and multilayered core-shell particles that have at least one magnetic component. The magnetic component enables control to drive self-assembly, while the nonmagnetic components can provide desired functionality, e.g., photonic, electronic, biological, etc. Moreover, the process is reversible in that the particles can be repeatedly assembled and redisperesed by applying and removing the bias field, thereby enabling the reversible creation of functional media on demand. Alternatively, assembled structures can be transferred intact to a substrate to form thin films. As such, the method opens up opportunities for the scalable bottom-up fabrication of nanostructured materials with unprecedented functionality for a broad range of applications. The computational model is very general and can be applied to many forms of self-assembly to provide a fundamental understanding of underlying mechanisms and enable the rational design of novel media.