The asymptotic performance of systematic rateless codes is first analysed using Gaussian approximation (GA) based on mutual information, which provides more accurate decoding thresholds than the method based on message mean. A modified linear program (LP) algorithm is proposed, where two key parameters are precomputed to efficiently search degree distributions with low overhead and appropriate average degree. The asymptotic analysis and simulation results show that the optimized code outperforms the codes obtained by conventional LP. Furthermore, the effects of outer code on overall code rate and decoding complexity are discussed.