In this paper, we consider partial sums of martingale differences weighted by random variables drawn uniformly on the sphere, and globally independent of the martingale differences. Combining Lindeberg's method and a series of arguments due to Bobkov, Chistyakov and Götze, we show that the Kolmogorov distance between the distribution of these weighted sums and the limiting Gaussian is "super-fast" of order (log n)^2 /n, under conditions allowing us to control the higher-order conditional moments of the martingale differences. We give an application of this result to the least squares estimator of the slope in the linear model with Gaussian design.