Use personas to get more specific voices

In our last example, we wanted GPT-3 to respond to a customer as a polite and helpful customer service agent. The use of the terms "polite" and "helpful" guide the model's response style, prompting it to respond as a certain persona. We can experiment with these terms to see how it affects the model's response. Let's change "polite, helpful" to "rude" and observe the model's response.

Asking GPT-3 to respond rudely will in fact prompt a rude response. Be careful what you wish for.

Specify the tone or persona you wish GPT-3 to emulate. Asking GPT-3 to respond rudely will result in a rude response. Be careful what you wish for.

GPT-3, being the amoral algorithm it is, accurately responds to the question, albeit vaguely, with a tone that better identifies with the word "rude". We can go as far as asking the model to respond as famous people or popular fictional characters. This can be used to our advantage as a shorthand to create a particular tone for the response.

---