Observational methods have been applied to assess work-related musculoskeletal disorders (WMSDs). The reliability of their results is a challenge. WMSDs remain a significant concern for workers, emphasizing the need for reliable observation tools. This study aimed to investigate the inter-rater reliability (IRR) of the Rapid Upper Limb Assessment (RULA) in assessing the posture of sculptors performing three tasks (sifting, kneading, and modeling). A fully crossed random effects model was applied to analyze IRR across thirty-nine subjects, three tasks, and three raters. Each rater individually evaluated a total of 2,340 postures. The IRR for each pair of raters was calculated using the weighted Cohen’s Kappa statistic. The highest IRR values among the rater pairs were found in the kneading task when assessing the trunk (1–2: IRR = 0.875; 95% CI [0.849–0.900], 1–3: IRR = 0.887; 95% CI [0.878–0.896], and 2–3: IRR = 0.934; 95% CI [0.923–0.945]). The worst IRRs were obtained for the wrist assessment (1–2: IRR = 0.168; 95% CI [0.159–0.177], 1–3: IRR = 0.145; 95% CI [0.138–0.152], and 2–3: IRR = 0.158; 95% CI [0.148–0.168]). The IRRs differed among tasks. Raters demonstrated higher IRR when scored upper arm, trunk, and leg postures, but fair reliability was observed for neck and wrist postures. Raters showed better IRR values when assessing postures for the sifting and modeling tasks than for the kneading task. This underscores the need for proper and adequate training before using RULA. In addition, task complexity seems to be an important aspect of IRR in this study. Further studies are needed to analyze how reliability is affected when analyzing different tasks.