The Dempster-Shafer theory of belief functions [Shafer 1976] is an intuitively appealing formalism for reasoning under uncertainty. Several AI implementations have been undertaken [e.g., Lowrance et al. 1986, Biswas and Anand 1987], but the computational complexity of Dempster's rule has limited the usefulness of such implementations. With the advent of efficient propagation schemes in Markov trees [Shafer et al. 1987], the time is ripe for more powerful systems. This paper discusses DELIEF (Design of bELIEFs), an interactive system that allows the design of belief function arguments via a simple graphical interface. The user of DELIEF constructs a graph, with nodes representing variables and edges representing relations among variables. This graph serves as a default knowledge schema. The user enters belief functions representing evidence pertinent to the individual variables in a specific situation, and the system combines them to.obtain beliefs on all variables. The schema may be revised and reevaluated until the user is satisfied with the result. The Markov tree used for belief propagation is displayed on demand. The system handles Bayesian causal trees [Pearl 1986] as a special case, and it has a special user interface for this case.