Flow
For my Master’s Capstone project with Microsoft, I led the end-to-end design of Flow, a neurodivergent-friendly AI assistant and toolkit that rethinks how large language models support neurodivergent workers.
From generative research to prototyping to testing, I designed 3 key features: Magic Mouse, Activity History, and a Custom AI System Prompt. These features work together to reduce context switching, ease communication stress, and scaffold focus. Flow transforms AI from a passive tool into a proactive partner, setting a new direction for inclusive, future-forward productivity design.
Sponsor
Microsoft Azure AI
Timeline
6 Months (Feb - Aug 2025)
Role
Product Designer
Impact
"This is such a real problem affecting many AI applications today. Love the focus on reducing cognitive load and it following you wherever you go! This is really cool to see!"
Senior PM @ Microsoft
"Wow, I'm looking at this and I feel like I want something like this. When I talk to my manager, this would be super helpful for me personally."
Senior Product Designer @ Microsoft
Adoption & Usability
100%
all 4 usability test participants stated that Flow would improve their workflow, and expressed interest in adopting it.
Time Saved
25 mins
users context switch less and stay in flow for longer, saving a projected 25 mins per day.
Company Recognition
4+
Microsoft Designers and PMs acknowledged Flow's value and potential impact.
AI Prompt Satisfaction
4.3/5
25 surveyed participants rated their satisfaction with the AI system prompt an average of 4.3 out of 5.
Problem Discovery
Context
Microsoft wanted us to leverage LLM to design an AI-powered solution for individuals with ADHD. We scoped our focus to neurodivergent tech workers, specifically Product Managers, Software Engineers and Designers, all of whom experience high-demand environments that can amplify ADHD pain points.
Access
Microsoft's sponsorship gave us access to their employees for research & testing.
Scalability & Business Impact
if our concept worked in this context, it could scale to other knowledge workers, which aligns with Microsoft’s goals of supporting all employees and shaping productivity tools.
Mixed Methods Study
To understand the problem space, we conducted desk and generative research. Given the sensitive and medical nature of ADHD, we created a study plan, recruited and conducted a mixed-methods study to learn from both lived experiences and scientific expertise.
Insight 1
Professionals with ADHD often feel emotional overwhelm or rejection sensitivity during social interactions. They need a judgment-free space to pause, rehearse, and regain clarity and communicate confidently.
Insight 2
Because ADHD causes fluctuating focus and energy levels, professionals need flexible systems that adapt to changing mental states, unlike rigid tools that can create additional friction.
Insight 3
Professionals with ADHD experience fragile momentum and benefit from support that minimize disruptions, protect focus, and helps them quickly recover context so they can resume tasks without losing flow.
How Might We
Through our research, participants emphasized that they did not view ADHD as a liability. Instead, they described it as a source of unique strengths such as hyperfocus, non-linear thinking, and authenticity. At the same time, they highlighted recurring barriers that made it difficult to fully leverage these strengths. The most consistent pain points were context switching, overwhelm, and workplace communication stress.
The challenge: Design a concept that amplifies these 3 strengths while reducing the cognitive and emotional friction that holds them back.
Final Design
Magic Mouse
Magic Mouse reduces context switching and cognitive load by letting users capture, query, and act on information directly in their workflow: no copy-paste or app-switching required. It also dynamically shifts between "text highlight" or "area select" mode for adaptive, context-aware support.
Magic Mouse: Text Highlight mode
Where longer, multi-page text or precision is required.
Magic Mouse: Area Select mode
Where non-text assets (e.g. images) are involved.
ADHD-Friendly System Prompt
I led the design of the custom AI system prompt that powers Flow's AI Assistant, ensuring neurodivergent-friendly communications. It provides the substance behind the chat window that appears after users invoke Magic Mouse.
Through multiple rounds of user and subject-matter-expert testing, I iterated on the tone, response structures, and guardrails to make interactions proactive, supportive, and low-friction.
Activity History
Activity History passively captures users work sessions through screen recording. It then creates a dynamic, searchable timeline that helps users quickly revisit and recover from interruptions, sustaining their flow and momentum.
Design Overview
2-Layered Approach
Flow needed to be designed at two layers because neurodivergent pain points show up both in how users access support and in how support is delivered.
Design Guidelines
Through the research process, I paid close attention to things that participants explicitly said that they did and did not want. I also synthesized overarching design guidelines that I wanted our concept to embody.
Be quietly available, never intrusive
show up when needed, stay invisible when not. Support should reduce cognitive load, not distract by spotlighting.
Proactive & Frictionless
anticipate moments of overwhelm or “blank canvas” paralysis. Provide simple, concrete starts that reduce effort and minimize context switching.
Emotionally Safe
Validate feelings without labels or judgment. Reframe perspectives to create a space where users feel supported rather than blamed.
Adaptive & Personalized
design for flexibility, adapting to different needs and coping strategies. Support should always feels relevant and never one-size-fits-all.
Interface Design
UI Layer
Process
We ideated widely, creating and refining user personas and information architecture. Through rapid sprints, design critiques with Microsoft Designers and University Professors, and user testing, we converged on a design.
How many features?
User testing and design critiques with Microsoft Product Designers revealed that less is more. Focusing on the features users valued most, enabled us to streamline the design and optimize each features impact.
Toolbar function & placement
The toolbar served as Flow’s entry point, giving users access to features and the AI chat. I experimented with different placements and levels of progressive disclosure to balance 2 competing needs: providing enough context for users to understand Flow's features, while keeping the toolbar lightweight and non-intrusive.
Early iterations tested expandable toolbars which surfaced features either one at a time or in a summary view. While this offered flexibility, I realized that progressive disclosure introduced additional cognitive load: users had to click, expand, and scan before accessing what they needed. The final version is minimal and non-collapsible, centered at the bottom of the screen. Its consistent placement made it easy to find, while its compact size kept the workspace clear. This made it feel more like a supportive layer rather than a dominant application, accessible but not overwhelming.
Iteration 1
expandable toolbar at the top-right of screen; showing 1 feature at at time in expanded view
Iteration 2
expandable toolbar at bottom-right of screen; showing summary of all features in collapsed view
OS layer vs. App
We first explored an OS-layer integration, which would have allowed Flow to feel ever-present, running seamlessly in the background like a built-in feature. However, this approach would have excluded anyone not running the latest version of Windows, creating a significant barrier to adoption. We chose instead to position Flow as a Microsoft Suite app. While this meant Flow wouldn’t have the “always-on” presence of an OS feature, it ensured accessibility across platforms, easy integration with tools like Teams and Word, and scalability for both Windows and non-Windows users.
Prompt Design
AI Layer
I led the design of Flow's interaction blueprint, designed on LLM GPT-4o, deployed via Microsoft Azure AI Foundry. I designed it to be modular and formatted with numbers (instead of a single chunk of instructions) to make it easy to make precise and targeted iterations post-testing.
Why Necessary?
Our research revealed that 56.4% of participants struggle with workplace communication, often to the extent of avoiding responses when feeling emotionally overloaded. This is consistent with our Insight 1: "Navigating Social Communication", gleaned from our research. Hence, they need additional communication-scaffolding through tailored responses.
😖 Over-Analyze
😪 Emotional Burden & Effort
Process
Round 1
Internal Evaluation
Each team member created a system prompt variation based on generative research, tested across the same 3 scenarios, and combined the best-performing elements into an improved prompt.
Round 2
ADHD User Evaluation
Ran semi-structured interviews where ADHD participants prompted the AI Assistant across 3 scenarios, sharing their expectations and real-time reactions.
Round 3
External Evaluation
Used Azure AI Foundry to conduct a likert-scale evaluation, ran 25 user surveys, and conducted semi-structured interviews with a psychiatrist and an ADHD coach to ensure the responses were clinically appropriate and aligned with real-world practices.
How it differs from Standard GPT?
Comparing the response to the same input, our customGPT is proactive, supportive and low-friction. Designed for neurodivergent users, it prioritizes psychological safety, empathy and actionable steps.
Standard GPT-4o
Flow's Custom GPT
Hand Off
Stakeholder Presentation
To hand off, we gave a final presentation to our sponsor and Master’s program professors. We also delivered design artifacts, including high-fidelity prototypes, the system prompt, and research documentation.
Next Steps
Our Microsoft sponsor told us that they will adapt and implement Flow’s designs, using our deliverables as a foundation for future iterations.
Impact Recap
"This is such a real problem affecting many AI applications today. Love the focus on reducing cognitive load and it following you wherever you go! This is really cool to see!"
Senior PM @ Microsoft
"Wow, I'm looking at this and I feel like I want something like this. When I talk to my manager, this would be super helpful for me personally."
Senior Product Designer @ Microsoft
Adoption & Usability
100%
all 4 usability test participants stated that Flow would improve their workflow, and expressed interest in adopting it.
Time Saved
25 mins
users context switch less and stay in flow for longer, saving a projected 25 mins per day.
Company Recognition
4+
Microsoft Designers and PMs acknowledged Flow's value and potential impact.
AI Prompt Satisfaction
4.3/5
25 surveyed participants rated their satisfaction with the AI system prompt an average of 4.3 out of 5.
Product Demo (4 mins)
Scripted, acted, filmed and edited by the team to showcase Flow in action