LLM hallucinations refer to the output of large language models that is not coherent with observable or known truth, but are actually mistakes made by the model trying to predict the most likely outcome. These "hallucinations" should be referred to as mistakes, and it's essential to focus on minimizing the error rate, just like for humans. To build resilient LLM-powered systems, strategies such as adding more context, using multi-head supervision, labeling output, applying output constraints, and specializing models can be employed. These approaches are not exclusive and work best in combinations, and they share similarities with traditional methods used for teams of humans or other software systems. By designing for errors and implementing controls, fault-tolerant approaches, and continuous improvement, organizations can unlock the benefits of LLMs while minimizing their mistakes.