Wireless video streaming often suffers from heavy and varying amount of packet loss. Short-length Luby Transform (SLLT) codes are often employed to offer necessary rateless or universal erasure protection. Nevertheless, a proficient method for finding the optimal degree distributions of SLLT codes is still missing although the asymptotic behaviors of infinite-length LT codes have long been deduced. In this paper, we propose a practical approach to find the optimal degree distributions of LT codes with arbitrary input block lengths. Our approach starts with a new fitness model for the LT codes based on three performance measurements: their coding overhead ε, decoding failure ratio r and failure probability p. We then employed the exponential natural evolution strategy (xNES) along with a suitable choice of initial degree distribution and a proper transformation between the genotypes and the phenotypes to search for the optimized degree distributions. The optimized SLLT codes designed using this approach far outperforms the ones obtained through asymptotic approximation. As an example application, we constructed a family of rateless convolutional UEP codes using these optimized SLLT codes as the rateless postcodes. These codes offer superb universal unequal erasure protection to wireless H.264/SVC broadcasting with only 16%-20% coding overhead. This design exercise, however, made clear the necessity of matching the rateless erasure correction capability of the SLLT postcode with the error-rate sensitive UEP capability of the convolutional precode. The effect of code matching was demonstrated clearly in the improvement of rate-distortion performance of video playback.