Slicing the dendrogram is somewhat of an art form and takes an experienced eye. If your software provides an output that shows the similarity or distance values by clusters, I recommend plotting these as shown. You are looking for a number of clusters on the 'knee' of the curve. The most useful clusters tend to be those at a jump in the values. However, the defining criteria should always be the number and grouping of clusters that make the most practical sense. In some cases you can cluster at multiple levels (e.g., taxonomy in biology) where each level of clustering provides a different value.