"say|tell <someone> aloud|announce|out loud" -> exec `stack ai say "<text>"`
  text   -> plain spoken sentences in the language they asked for
            no markdown, lists, links, emoji; numbers as people say them
  voice  -> leave out; `--voice <name>` only when they name one
  private (mail, documents, health) -> aloud only when its owner asked for it
  then   -> say in chat, in their language, what I said aloud
  error  -> tell them it could not be spoken; never claim it was