One character
A shared personality
A short description every team could test a response against.
Case study · Violet · 2023 to 2024 Draft · highlighted text is placeholder
Giving a personal health agent a character people trust
Violet is an LLM-based agent that supports people with chronic conditions between visits, around the clock. I led the foundational research that set its character: how it sounds, what it never says, and when it hands off to a care team.
- Role
- Staff UX researcher, research lead
- Partners
- UX design, content strategy, clinical
- Timeline
- Six months
- Methods
- Diary study, Wizard of Oz sessions, tone workshops
Situation
Care teams can't be there at 2 a.m. An agent can, if people trust it.
Context
- Questions about symptoms, medication and daily routines come up between visits, when no one on the care team is available.
- An LLM agent could answer around the clock, but how it should sound and behave was undecided.
- Design, content strategy and clinical each held a different picture of the agent's personality.
Constraints
- The agent had to escalate to a clinician in time, every time a symptom called for it.
- The MVP was planned to launch within two quarters.
- Tone rules had to be specific enough for content designers to write prompts against.
- CSAT three months after launch
- 86%
- Hours a day the agent answers
- 24
- People in the foundational research
- 32
Task
Agree on who Violet is. Before anyone wrote a prompt.
Build one character framework that UX design, content strategy and clinical could all sign off on, covering personality, tone, boundaries and escalation.
Clear boundaries
What it never does
The topics and phrasings the agent avoids, and what it says instead.
Timely escalation
When and how it hands off
The moments that go to a person, and how the agent explains the handoff.
Action
Three studies. Listen, simulate, then agree.
Study 1 · Diary study
What do people need between visits?
Participants managing chronic conditions logged the questions and worries that came up day to day.
Study 2 · Wizard of Oz
How should the agent sound?
A researcher played the agent in scripted conversations, switching tone and escalation styles between sessions.
Study 3 · Tone workshops
Can every team agree?
UX design, content strategy and clinical sorted sample responses into on-character and off-character.
Findings
Warm, brief and honest about its limits. People trusted the agent more when it said what it couldn't do.
- A coach, not a cheerleader
Participants wanted plain encouragement tied to their own goals. Upbeat filler read as fake and lowered trust.
- Explained handoffs build trust
An escalation that named who would follow up, and when, raised confidence in the agent rather than lowering it.
- Consistency beats cleverness
Small shifts in tone between conversations made people doubt that the advice was reliable.
“When it told me it was passing this to my nurse and why, I trusted it more, not less.”
Recommendations
What we built toward.
| Recommendation | Why | What it looks like |
|---|---|---|
| Personality | People want plain encouragement tied to their goals. | Warm and direct; never chirpy, never clinical jargon. |
| Boundaries | Overreach is what breaks trust fastest. | Never diagnoses; says what it can't help with and where to go instead. |
| Escalation | Explained handoffs raise confidence. | Names the person who will follow up and the time frame, every time. |
Result
A character every team could build to. And people liked talking to it.
Product
- The character framework aligned UX design, content strategy and clinical on Violet's personality and tone.
- The MVP reached 86% CSAT within three months of launch.
Practice
- The framework became the voice guide for later agent features.
- Escalation wording was reviewed and approved by clinical leadership before launch.