@mttaggart@infosec.exchange it actually makes a lot of sense that a Large Language Model would use the tone of voice and other features of human speech in order to tell the difference between the roles instead of using the XML style tagging which don’t look like human language.