We introduce Backward Conformal Prediction, a new method that adapts coverage levels to enforce interpretable, data-dependent prediction set sizes, with provable guarantees and practical estimators.