Background: Repetition of hospital-treated deliberate self-harm is common. Several recent studies have used emergency department data to develop clinical tools to assess risk of self-harm or suicide. Longitudinal, linked inpatient data is an alternative source of information. Methods: We identified all individuals admitted to hospital for deliberate self-harm in two Australian states (~350 hospitals). The outcome of interest was a repeated episode of self-harm (non-fatal or fatal) within 6 months. Logistic regression was used to identify a set of predictors of repetition. A risk calculator (RESH: Repeated Episodes of Self-Harm) was derived directly from model coefficients. Results: There were 84,659 episodes of self-harm during the study period. Four variables – number of prior episodes, time between episodes, prior psychiatric diagnoses and recent psychiatric hospital stay – strongly predicted repetition. The RESH score showed good discrimination (AUC=0.75) and had high specificity. Patients with scores of 0–3 had 14% risk of repeat episodes, whereas patients with scores of 20–25 had over 80% risk. We identified five thresholds where the RESH score could be used for prioritising interventions. Limitations: The trade-off of a highly specific test is that the instrument has poor sensitivity. As a consequence, the RESH score cannot be used reliably for “ruling out” those who score below the thresholds. Conclusions: The RESH score could be useful for prioritising patients to interventions to reduce readmission for deliberate self-harm. The five thresholds, representing the continuum from low to high risk, enable a stepped care model of overlapping or sequential interventions to be deployed to patients at risk of self-harm.