In the Camunda modeler, when I use English words as userTask names, everything is ok.
But when I use other language like Persian which is in RTL mode, when words are more than three words, words arrange is messed up.
we know about problems regarding RTL languages inside the Camunda Modeler . It’s not an easy task to support it in our modeling libraries. However, there are some examples  to modify the Camunda Modeler to kind of deal with it. In this case, you will have to modify the underlying bpmn-js library which is used in the Modeler. We are open for contributions (e.g. a plugin which fixes this)
Mind sharing a screenshot of what it looks like?