Mind Keep: Mental Health CUI

Overview

Mind Keep emerged from a practical accessibility problem: many people want help processing stress, anxiety, or negative thought patterns, yet traditional care is often expensive, delayed, logistically difficult, or emotionally difficult to initiate. I framed the project as a supplementary support tool rather than a replacement for therapy, then designed a conversational system that could help users reflect, regulate, and organize their thinking in the moment.

I focused the concept on three core tasks. First, structured CBT sessions for identifying thoughts, emotions, and distortions before reframing them. Second, guided meditations and breathing exercises for immediate regulation. Third, lightweight mental-health Q&A for people who needed quick support without entering a formal exercise.

Problem Space

The source material pointed me to three access barriers: financial, logistical, and psychological. My report cites therapy costs in the $125 to $250 range per 50-minute private session, while Mental Health America data showed a sizable share of adults with extended mentally unhealthy periods still could not access care due to cost. Even when people want support, the surrounding system can still require waitlists, forms, scheduling friction, and willingness to disclose personal struggles to another person.

That gap creates space for a different point of entry. A conversational system can be available 24/7, function on phone or desktop, and reduce the commitment barrier that causes many people to avoid formal care. The central design challenge was not availability alone, but designing a system structured enough to be useful, empathetic enough to feel safe, and appropriately constrained so that it would not overstep.

Goals

Research

I grounded the concept in both literature and real therapeutic structure. In my final report, I drew on APA descriptions of CBT, systematic review findings across Woebot, Wysa, and Youper, and research on authenticity limits in AI empathy. That combination was important because it pushed the design away from generic chatbot behavior and toward a repeatable therapeutic scaffold.

I also used my midterm work to examine real CBT transcripts and break them into reusable interaction components such as open inquiry, situation understanding, challenge, reflection, and homework. That move let me translate clinical flow into interaction architecture instead of copying therapy aesthetics.

Personas + Early Concepts

I strengthened the concept in the midterm phase by designing for two distinct usage modes instead of one generic "mental health user." Lauren represented the structured CBT path. Alex represented quick de-escalation and recovery through guided meditation.

These personas directly shaped my early testing methods. For the CBT flow, I used a Think Aloud paper prototype because the dialogue needed to stay reactive to whatever the user wrote. For the meditation flow, I used a Wizard of Oz format with text, audio cues, and simulated guidance because the real value came from how the interaction felt while the user disengaged from the screen.

Solution + Prototype

I shifted the final prototype from paper to a working custom GPT. That let me test real dynamic responses instead of static branching. It also let me encode behavioral rules at the instruction level: tone, scope, crisis handling, consent checks, pacing, and when to offer structured exercises.

I also made a pragmatic platform choice. I considered BotPress, LandBot, and VoiceFlow, but selected ChatGPT because the responses were more reactive and substantially less constrained than rigid pre-authored flows. That mattered in this CUI because the quality of therapeutic pacing depended on the system's ability to follow a user's actual context.

Testing + Findings

I used two useful testing phases in this project. Early on, I used focus-group-style questions, paper prototyping, and Wizard of Oz simulation to understand whether the concept felt approachable at all. In the final round, I tested the custom GPT with two pilot users and an exploratory review from a psychology doctoral student.

A consistent pattern I found across both the midterm and final materials was that people responded more strongly to structure than personality. That insight helped me understand that the product performs best when it acts as a clear guide with a supportive tone, rather than when it attempts to simulate human closeness too aggressively.

Safety + Limitations

Because I was designing for mental health support rather than a typical productivity or commerce flow, I treated the safety decisions as first-class design work. My reports repeatedly return to four constraints: transparency, crisis escalation, privacy, and longitudinal memory.

In my final report, I also identified engagement as an important product risk. Even if people value the tool, long-term use is not guaranteed. That means success should not be measured only by immediate satisfaction. It should also be measured by whether the product can maintain trust, adapt response depth, and create a reason to return without becoming repetitive.

Next Steps