Clinicians have a new tool, NBvarsel, that chooses which clinical variables to include in a prediction model by maximizing net benefit. The algorithm compares every possible set of predictors, each evaluated with cross‑validation to compute net benefit. For small sets it runs an exhaustive search. When many predictors are available, it uses a backward‑elimination version. In three real studies – ovarian‑cancer risk, emergency‑department hospitalisation and non‑typhoidal Salmonella bloodstream infection – NBvarsel produced short, clinically useful models. The new models matched or outperformed those built with standard backward elimination (AIC) and LASSO regularisation in discrimination, calibration and net benefit. The exhaustive search can be computationally heavy; with twenty candidates it would fit over five million models with five‑fold cross‑validation. Therefore, the method is best for datasets with a limited number of predictors or when an earlier variable filtering step is used. Assigning a test‑harm value for each variable is subjective and may vary by setting. If harms cannot be set, the algorithm still works but may see smaller gains over conventional methods. The authors recommend using it only when the sample size satisfies Riley’s guideline and after any needed a‑priori variable reduction. The package implementation is available in R and the full code can be accessed via the authors’ GitHub repository.