I’m working with a faculty member who wanted an AI persona for practicing difficult workplace conversations. The end goal is audio interaction and maybe a visual representation.
I try to approach projects like these in stages. Can I build the stages out so that if we hit a wall, the stuff we’ve built to that point is still useful?
Given that accurate responses from the persona is the foundation of the experience, we need to figure out what elements are needed to create a persona that behaves (mostly) as expected. Having a few standard scenarios and interaction patterns to apply to the provisioned personas should help us see if the persona is working. I found making them to be tedious and it was hard to isolate variables. So . . . I built the small proof of concept persona prompt generator below.
In this case, we’re still determining what elements are needed to create a good persona- that’s some mix of attributes and the language nuance. I think the example is decent (but could use a few tweaks and I’m no expert). It is entirely focused on Hofstede’s Cultural Dimensions Theory so far. We can add some components like age or gender or contextual backstory, etc. That should make it easier to create additional personas to test against. It could even be an aspect of student work in a current class (what attributes create good personas and evaluating persona responses according to expected behavior).
Once we feel good about the personas, I can start working on the API options to move the persona seamlessly into ChatGPT. After that’s working well, we can start to look at audio interaction options and other more advanced features.