Motivated by the use of location-scale regression models to support several recent extremal regression methods, we consider the estimation of, and inference about, conditional extreme quantiles and related quantities based on standardized residuals obtained following a preliminary model estimation step. We show that residual-based versions of extreme value estimators are asymptotically normal, just as their unachievable counterparts based on unobserved regression errors would be, under a high-level condition which essentially requires the number of residuals that do not correctly predict the unobserved errors to be asymptotically smaller than the multiplicative inverse of the rate of convergence of the extreme value procedure. This condition is shown to be substantially weaker than corresponding conditions obtained in related recent work, and we discuss how our theory applies to a wide range of examples containing location-scale parametric regression models and autoregressive time series.