Crop breeding programs are interested in using genetic resources but have difficulty identifying useful accessions from germplasm collections. To efficiently use the diversity present in large germplasm collections, breeders often identify a subset of accessions that represent the larger collection. Methods to identify these subsets, which are called core collections, do not consistently capture functional diversity, and breeders would benefit from methods that help create custom core collections using existing data from variety trials or breeding programs. Making use of high‐density genomic data and existing phenotypic data from a collection of 433 domesticated carrot (Daucus carota L.) accessions, we tested whether it is possible to develop custom subsets of accessions for specific breeding purposes. We found that for this collection, representative strategies were effective in developing core collections that capture the diversity of the collection, but they were no better than random sampling, likely because the collection itself is not strongly subdivided. Custom strategies generated subsets that differed from the total collection with altered genetic, geographic, and phenotypic compositions. When used as training populations for genomic prediction of the other accessions in the collection, however, these custom cores did not produce a substantial improvement over traditional core collections. Increasing the size of the core did improve prediction accuracy, suggesting that it is possible to improve the usefulness of core collections by identifying custom subsets that are large enough to represent the functional genetic diversity present in the collection.