Natural language is an intuitive interface to supplement programming in modern automation settings. However, most natural language frameworks are specialized and not universally applicable. We contribute a novel layered pipeline that transforms the instructions of laypersons into robot programs with non-linear control flow and that facilitates reuse. The instructions are analyzed regarding grammatical features. From this, the syntactical analysis derives programs with nested control structures and references to physical parts within a scene to be manipulated. These programs are semantically interpreted during online execution in a concrete scene—i.e., the control structures are evaluated, and part specifications are grounded to physical parts. With that, a fully specified skill is created and executed by a robot system. Since only the input/output interface of each pipeline stage is defined, they are adjustable independently of each other. Our experiments demonstrate how industrial robots in diverse domains can be verbally programmed using the pipeline.