Multi-user transmission at 60 GHz promises to increase the throughput of next-generation WLANs via both analog and digital beamforming. To maximize the capacity, analog beams need to be jointly configured with user selection and digital weights; however, joint maximization requires prohibitively large training and feedback overhead. In this paper, we scale multi-user 60-GHz WLAN throughput via design of a low-complexity structure for decoupling beam steering and user selection such that analog beam training precedes user selection. We introduce a two-class framework comprising: 1) single-shot selection of users by minimizing overlap of their idealized beam patterns obtained from analog training and 2) interference-aware incremental addition of users via sequential training to better predict inter-user interference. We implement a programmable testbed using software-defined radios and commercial 60-GHz transceivers and conduct over-the-air measurements to collect channel traces for different indoor WLAN deployments. Measurements are conducted using a 12-element phased antenna array as well as horn antennas with different directivity gains to evaluate the performance of practical 60-GHz systems. Using trace-based emulations and high resolution 60-GHz channel models, we show that our decoupling structure experiences less than 5% performance loss compared with maximum achievable rates via joint user-beam selection.