Melanoma is one of the deadliest skin cancers. Recently, developed single-cell sequencing has revealed fresh insights into melanoma. Cytokine signaling in the immune system is crucial for tumor development in melanoma. To evaluate melanoma patient diagnosis and treatment, the prediction value of cytokine signaling in immune-related genes (CSIRGs) is needed. In this study, the machine learning method of least absolute selection and shrinkage operator (LASSO) regression was used to establish a CSIRG prognostic signature of melanoma at the single-cell level. We discovered a 5-CSIRG signature that was substantially related to the overall survival of melanoma patients. We also constructed a nomogram that combined CSIRGs and clinical features. Overall survival of melanoma patients can be consistently predicted with good performance as well as accuracy by both the 5-CSIRG signature and nomograms. We compared the melanoma patients in the CSIRG high- and low-risk groups in terms of tumor mutation burden, infiltration of the immune system, and gene enrichment. High CSIRG-risk patients had a lower tumor mutational burden than low CSIRG-risk patients. The CSIRG high-risk patients had a higher infiltration of monocytes. Signaling pathways including oxidative phosphorylation, DNA replication, and aminoacyl tRNA biosynthesis were enriched in the high-risk group. For the first time, we constructed and validated a machine-learning model by single-cell RNA-sequencing datasets that have the potential to be a novel treatment target and might serve as a prognostic biomarker panel for melanoma. The 5-CSIRG signature may assist in predicting melanoma patient prognosis, biological characteristics, and appropriate therapy.