Visual information received from the three types of photoreceptor neurons(R1-R6, R7 and R8) in the fly compound eyes converges to the external part of the medulla neuropil (M1-M6 layers) in a layer-specific fashion: R7 and R8 axons terminate at the M6 and M3 layers, respectively, whereas lamina neurons(L1-L5) relay R1-R6 to multiple medulla layers (M1-M5). Here, we show that during development, R7 and R8 neurons establish layer-specific projections in two separate stages: during the first stage, R7 and R8 axons sequentially target to the R7- and R8-temporary layers, respectively; and at the second stage, R7 and R8 growth cones progress synchronously to their destined layers. Using a set of mutations that delete different afferent subsets or alter R7 connectivity, we defined the mechanism of layer selection. We observed that R8, R7 and L1-L5 afferents target to their temporary layers independently,suggesting that afferent-target, but not afferent-afferent, interactions dictate the targeting specificity. N-cadherin is required in the first stage for R7 growth cones to reach and remain in the R7-temporary layer. The Ncad gene contains three pairs of alternatively spliced exons and encodes 12 isoforms. However, expressing a single Ncad isoform in Ncad mutant R7s is sufficient to rescue mistargeting phenotypes. Furthermore, Ncad isoforms mediate promiscuous heterophilic interactions in an in vitro cell-aggregation assay. We propose that Ncad isoforms do not form an adhesion code; rather, they provide permissive adhesion between R7 growth cones and their temporary targets.