BUILDING MINDSHIFT: DESIGNING AN AI GRIEF COMPANION FOR ADOLESCENTS
What happens when you try to build a mental health app for teenagers? We submitted MindShift to the MIT Global Appathon. Here is what we built, what we got wrong, and what the design decisions actually cost us.
MindShift started from a simple observation: grief resources for adolescents are mostly designed for adults or very young children. There is almost nothing for 10-17 year olds who understand what loss means but lack the emotional vocabulary or professional access that adults have.
I built MindShift with Aroush Muglikar and submitted it to the HackerBoundary Global AI Hackathon.
Mira — The AI Companion
Mira is a custom-prompted conversational AI designed specifically for grief support with adolescents. The prompt engineering was the hardest part. General-purpose AI companions either over-pathologise (treating every sad message as a crisis) or under-respond (generic positive affirmations). We designed Mira around three principles: validate first (never jump to problem-solving), use age-appropriate language (no clinical jargon), and know when to suggest professional help (specific triggers identified from adolescent mental health literature). The system prompt is about 800 tokens and went through around 20 iterations before the tone felt right.
The Grief Journey Tracker
Grief is not linear — the Kübler-Ross stages model has been largely abandoned in clinical practice. What is more useful is letting adolescents visualise their emotional state over time without telling them what they "should" be feeling. The tracker is a daily check-in mapping emotional state to a visual timeline. No scores, no progress bars — just a visual record. Seeing your own emotional history is more useful than comparing to a normalised model.
The Breathing Engine
Physiological regulation before emotional processing — this is evidence-based. If someone is acutely distressed, cognitive-level interventions do not work well until the nervous system is regulated. The breathing engine implements box breathing (4-4-4-4) and 4-7-8 breathing with animated visual guidance.
The Memory Gallery
Designed for grief specifically, not general mental health. The ability to hold positive memories of someone who has died is central to healthy grief processing — what psychologists call "continuing bonds theory." Users add photos, text, and voice notes organised around a person they have lost. Mira can reference the gallery in conversations.
The Teacher Note Generator
Teachers often do not know how to talk to students who are grieving. The generator produces templated, age-appropriate notes from the student's perspective that explain what kind of support would help. This was the most practically useful feature we built.
What We Got Wrong
The emotional analytics dashboard was a mistake. It created a kind of performance pressure — people feeling like they should be "improving" on their grief metrics. That is exactly the wrong mental model. We would remove it in v2. The glassmorphism UI also creates contrast issues on lower-quality screens. Mental health tools need to be accessible by definition.
What Building This Taught Me
Domain-specific AI prompting is genuinely hard. General-purpose prompts produce general-purpose responses — which in a grief support context can feel cold, clinical, or occasionally wrong. The specificity of the system prompt matters enormously. Getting the tone right for a 13-year-old who just lost a parent took more iteration than the entire frontend. The intersection of AI and mental health is one of the most consequential areas in software right now, and getting it wrong has real costs.