While exoskeletons have markedly reduced the metabolic cost of walking, exoskeleton design remains expensive and time-consuming. Biomechanical simulations could improve exoskeleton development with their ability to optimize many design and control parameters simultaneously. While promising, simulations often rely on simplifying assumptions, such as fixed kinematic methods. Furthermore, few simulation-based designs have been tested directly in human experiments. To evaluate the utility of simulations for exoskeleton design, we designed a controller using biomechanical simulation and then tested the controller with a hip-knee-ankle exoskeleton worn by a single experienced user. Here we show that a biomechanical simulation constrained by safe, experiment-based torque limits can find an assistance pattern that reduces metabolic cost but with substantial differences between expected and measured outcomes. We found a simulated assistance profile that was expected to reduce the metabolic cost of walking by 69.0%. When tested experimentally, it reduced the cost by 25.9% compared to walking in the device unassisted. The simulation predicted the direction of activity change in most muscles but overestimated the total magnitude of these changes. The applied torques resulted in joint kinematic and stride frequency changes that were not accounted for in the simulation. Future simulations could allow kinematic adaptations and use updated cost functions that better reflect how humans respond to assistance. Continued feedback between experimental testing and simulation design will build on these promising results to improve the accuracy of simulations and enhance their ability to guide future exoskeleton designs.