Construction and running a dynamical ensemble only provides a starting point for estimating the distribution of the true future state. To account for the residual errors missing from the dynamical ensemble, enlightened by the best member dressing technique of Roulston and Smith (2003), we augment the dynamical ensemble in the post-processing by adding each member of the dynamical ensemble a certain number of statistical perturbations that are called the dressing perturbations. Different from the best member method where the dressing perturbation statistics comes from the archived historical best member error, the new dressing kernel is determined by making the dressed ensemble members indistinguishable from the verifications under the second moment measurements on the seasonally averaged basis. To test this new dressing technique and compare it with the best member method we run the ensemble transform Kalman filter (ETKF) ensemble. In the test categories of rank histogram, skill scores and ensemble covariance precision, the new dressing kernel performs better than the best member method. This study also reveals the problems and questions the hypothesis of the best member method.