STOA

AI-guided anxiety intervention, exactly when you need it.

Role

Solo Designer

Timeline

Winter 2026

Tools

Figma

Elicit

Claude Opus 4.6

Figma Console MCP

Platform

Smartphone

Smartwatch

Introduction

The AI-powered wearable that intercepts acute anxiety

01

Proprietary smart band detects elevated physiological stress before panic peaks

02

App delivers a CBT-grounded anchor — breathing, grounding, or a Stoic reframe — immediately

03

Pattern learning personalizes detection thresholds over time

Success Criteria
  • Intervene before cognitive function is fully impaired

  • Complete an anchoring technique in less than 2 minutes, in a public setting

  • Zero false-alarm fatigue: detection precise enough that every alert means something

Problem

Panic attacks don’t wait for you to ask for help

Anxiety disrupts the ability to think clearly and act, yet existing tools wait for the user to open an app and start an exercise at exactly the moment they are least able to. Most people never do; avoidance is the most common response.

Solution

I built something that doesn’t wait

STOA reads the body’s stress signals and starts an intervention on its own, cognitive, behavioral, or somatic, matched to the moment. It covers the gap between needing help and being able to seek it.

Video

Introducing STOA.

A quick walkthrough of the AI anxiety intervention: the smart band detection, the anchor flow, and the Stoic voice that carries it.

My Role

Solo designer

Responsibilities

User Research, Empathy Mapping, Jobs-to-be-Done, Design System, UI/UX Design, Prototyping, Wizard-of-Oz Testing

People

Just me + Claude as an AI design partner

Timeline

6 weeks from research to prototype

Research

Built on what the literature actually says

I used Elicit, an AI-powered research assistant, to systematically synthesize 42+ academic sources. What surfaced drove every design decision that followed.

01

Anxiety sensitivity is the real problem.

People fear the sensations of anxiety, the racing heart, the tight chest, more than whatever causes them.

02

Random delivery makes things worse.

Interventions triggered by personal physiological thresholds significantly outperform scheduled reminders or random notifications.

03

Wearables work.

Consumer smartwatches show strong accuracy for anxiety detection. Personalized machine-learning models outperform generic approaches after a 1–2 week learning period.

04

CBT is the framework. Stoicism is the voice.

Modern CBT descends directly from Stoic practice, so the Stoic voice is functional, not decorative: it's the original source of the cognitive techniques the app delivers. The evidence supports cognitive reframing alongside behavioral and somatic techniques, so STOA draws on all three, with Stoicism setting the tone.

How this shaped the design

They all pointed at the same requirement: the system has to make the first move, because someone mid-panic cannot be expected to open an app and search for help.

Process

An AI-assisted workflow from research to prototype

01

Literature Synthesis — 42+ sources synthesized with Elicit, across six research questions

02

Define — empathy map + Jobs-to-be-Done framework

03

CLAUDE.md Design Brief — visual and interaction rules, encoded once for every screen

04

Design System — tokens built to WCAG AAA contrast via Figma Console MCP

05

Storyboard & Screens — 10 AI-generated storyboard panels; UI built in Figma with Claude Code

06

Wizard-of-Oz + AI Critique — physical prototype tested live; screens critiqued by Claude as a skeptical engineer

Design

The design system built for a cognitive crisis

Every choice ran through one filter: is this usable for someone experiencing an acute anxiety attack, in public, with shaking hands? The answer was to strip the interface to almost nothing: two standout colors, Aegean Deep for calm and Bronze for action, and no red or anything else that could read as alarm. For typography, Plus Jakarta Sans for the interface and Cormorant Garamond for Stoic reflection. The color names — Aegean Deep, Bronze, Obsidian, Marble White — trace to the ancient Mediterranean, where Stoicism began.

System at a glance
109
Design Tokens
37 primitive, 24 semantic, 48 component
8
Type Styles
Cormorant for reflection, Jakarta Sans for action
4
Brand Colors
Calm tones for the body, deep tones for focus

Key Screens

Home Dashboard

One vital sign, one streak, one last session. A stressed user needs reassurance at a glance, so the dashboard holds everything else back.

Detection Alert

The bet the whole product rests on: detect an acute anxiety attack as it begins and reach out first, because no one opens a wellness app mid-panic.

AI Recommendation

The AI explains each recommendation, what it detected and why this technique fits, because an anxious user won’t act on advice they don’t understand.

Stoic Prompt

The words of Stoic philosophers instead of a clinical script. The prompt copy was written as deliberately as the layout was designed, because the message those words carry is what helps someone calm down.

Post Check-In

Five honest options, including “still anxious,” so the product records what actually happened instead of assuming success.

Progress

Episodes and recovery time, week by week, the numbers that show whether the product is actually helping.

Prototype

Try the prototype yourself

A live, clickable Figma prototype of STOA: move through the flow the way a user in an acute anxiety moment would, from first reading to recommended intervention.

Reflect

What designing for a crisis taught me about AI

01

The brief did more work than the model.

The CLAUDE.md design brief encoded the visual and interaction rules once, and every generated screen was held to it. Generating with AI is the easy part; the judgment is the work: deciding what’s true, what to cut, where the line sits. A sharp brief produces sharp output. A vague one produces confident nonsense.

02

An AI that can’t show its reasoning can’t be trusted, least of all mid-panic.

Designing the recommendation screen, the question was never which intervention the AI should pick, it was whether an anxious person would believe it. So every recommendation shows its evidence, the readings that triggered it and why this technique fits. Advice that arrives without a reason is easy to dismiss, and someone mid-panic will dismiss it.

03

AI is good at noticing. It’s bad at caring.

The watch detects the spike; the model proposes the intervention. But what actually calms someone — a Stoic voice instead of a clinical alert, the right tone at 140 BPM — came from research and empathy, not the model.

04

AI turned the job from making into directing.

Across research, copy, and the build itself, AI kept the raw material coming, and my job was choosing between it all: what to keep, what was off, what to cut. The moment I stopped steering, quality fell off.

What I’d do differently

01

Resolve the “Still anxious” path first.

I’d prototype the failure state before the success state. It carries the highest emotional stakes in a mental health product, and I left it unanswered.

02

Add consent before biometrics.

Onboarding jumps from anchor selection straight to a live BPM dashboard without explaining how physiological data is collected, stored, or used. A product that reads your body has to ask permission before it starts listening.

03

Test the error states.

The wearable band disconnecting mid-exercise. A failed session. A false detection. Each one lands on a user who is already on edge, and the prototype never answered them.

Next