In scientific domains, knowledge is often discovered from experiments by grouping or clustering them based on the similarity of their output. The causes of similarity are analyzed based on the input conditions characterizing a given type of output, i.e., a given cluster. This analysis helps in applications such as decision support in industry. Cluster representatives form at-a-glance depictions for such applications. Randomly selecting a set of conditions in a cluster as its representative is not sufficient since distinct combinations of inputs could lead to the same cluster. In this paper, an approach called DesCond is proposed to design semantics-preserving cluster representatives for scientific input conditions. We define a notion of distance for conditions to capture semantics based on the types of their attributes and their relative importance. Using this distance, methods of building candidate cluster representatives with different levels of detail are proposed. Candidates are compared using the DesCond Encoding proposed in this paper that assesses their complexity and information loss, given user interests. The candidate with the lowest encoding for each cluster is returned as its designed representative. DesCond is evaluated with real data from Materials Science. Evaluation with domain expert interviews and formal user surveys shows that designed representatives consistently outperform randomly selected ones and different candidates suit different users.