A system architecture is presented to generate sensor-controlled robot tasks from knowledge encoded in a CAD model. This architecture consists of an application layer where the user annotates assembly tasks in the CAD software. A process layer infers the specific robot skills and parameters from the CAD model and annotated data. A control layer executes the complex, force-controlled tasks. A proof-of-concept implementation is made, consisting of an application layer implemented in FreeCAD and a process layer that focuses on using fuzzy inference to generate appropriate skill-dependent process parameters from the geometric CAD information and annotations in the CAD model. In the control layer, a constraintbased control framework is used to robustly execute the assembly tasks. The system is validated on a challenging assembly task involving the assembly of screw compressor parts.