Community assembly provides the foundation for applications in biodiversity conservation, climate change, invasion ecology, restoration ecology, and synthetic ecology. Predicting and prioritizing community assembly outcomes remains challenging. We address this challenge via a mechanism-free LOVE (Learning Outcomes Via Experiments) approach suitable for cases where little data or knowledge exist: we carry out actions (randomly-sampled combinations of species additions), measure abundance outcomes, and then train a model to predict arbitrary outcomes of actions, or prioritize actions that would yield the most desirable outcomes. When trained on <100 randomly-selected actions, LOVE predicts outcomes with 2-5% error across datasets, and prioritizes actions for maximizing richness, maximizing abundance, or minimizing abundances of unwanted species, with 94-99% true positive rate and 12-83% true negative rate across tasks. LOVE complements existing approaches for community ecology by providing a foundation for additional mechanism-first study, and may help address numerous ecological applications.