Most iterative clustering algorithms require a good initialization to achieve accurate results. A new initialization procedure for all such algorithms is given that is exact when the data contain compact, separated clusters. Our examples use c-means clustering.