illustration of chat bubbles

Innovation: Using personas to prompt-engineer conversational AI Chatbots

Personas let us create compelling stories about our users that make their identity, needs, challenges, and goals tangible. These stores can be foundational to bringing empathy to experience design. I leveraged personas to shape a company-wide AI Hackathon activity–and ultimately seed an AI- assisted chatbot effort for HHS.gov.

User research

Site survey data revealed five primary audiences to the HHS.gov site and each audience group was looking for on the site.

One user group example is the “Agency Employee/Contractor” audience. Their most common goals were to:

  1. “Review policy language concerning trainings and conferences.”
  2. “Complete Cybersecurity Training to maintain access to the CCW VRDC.”
  3. “Learn more about recently updated (HHS) workplace flexibilities policy.”
  4. “Read the Updated (HHS) Workplace Flexibilities Policy.”

HHS USER PERSONAS

I converted our research insights into five distinct, relatable fictional characters. Each persona is a simplified representation of the key user groups; and each highlighted the needs, goals, and pain points of a fictional character.

For the purposes of the hackathon’s theme, we used AI to generate Clue game-themed player ‘cards’ for each user group. 

Persona

Screenshot: “Clue”-game styled persona card example

CHATBOT PILOT

We partnered with the development team to generate an AI chatbot pilot teams could use as a starting point. 

The prepackaged chatbot used Drupal’s AI Module and Open AI GPT for its Learned Language Modeling (LLM). We used Chat GPT and configured the module to enable retrieval-augmented generation (RAG) to deliver contextually-aware information unique to our target website: HHS.gov.

This approach supported my goal of having teams configure their chatbot(s) to respond uniquely to HHS.gov site visitors, depending upon their assigned personas. 

FROM PERSONAS TO PROMPTS 

In an authoring interface, each team could provide parameters for their chat agent, including custom instructions for how it should behave. Each team had the same starting point: a pilot chatbot and instructions on how to customize and launch it.

Screen for chatbot programmingActivity instructions

 

Screen for programmingUI for Chatbot programming

AI NEEDS HUMAN CENTERED DESIGN

Teams across the organization gained new perspectives on how human-centered design tools are integral to creating new AI initiatives.

Each produced chatbot revealed distinct differences between the chatbot’s tone, language use, and voice. Across all the persona-informed chatbots, we saw significantly more personalized communications than those delivered via the unprompted chatbot.

For federal agency websites, the enormity of information that’s generated and retained through federal research activities and communication artifacts is hard to fathom. It’s even harder to navigate efficiently.  Public communication and information access are the fundamental purposes of most federal agency sites.

Outcomes

 

I presented HHS client leadership our hackathon’s AI chatbots. Agency decision makers saw their potential value for improving how users find information on HHS.gov.

 

The feature is on the HHS.gov roadmap for Q4 2025 release.