Building a realistic training environment to teach situation awareness and coordination of teams implies more than good graphics and HCI. Plausible and varied behavior of the artificial characters, richness and variety of the stories experienced by the trainee are essential features of an effective training session. This paper discusses an authoring environment for simulations targeted at the training of firefighting commanders. This environment, called PRESTO-FF, currently includes a modelling facility for behaviours of people on the scene, a specific 3D environment, and a visual scripting system that drives the unfolding of the simulation according to the player's actions and real-time choices by a trainer. Models and scripts are reusable in multiple 3D environments thanks to abstractions and semantics provided by underlying AI technologies. Results of the validation of a simulation with domain experts are reported, highlighting both advantages and limitations of the current framework.