Mirror
Conversational AI Project
Conversational AI Project
Year
Year
2026
2026
Duration
Duration
7 weeks
7 weeks
Overview
Overview
Most students don't choose the wrong major because they lack information, but because they lack self-knowledge, or because they're quietly performing someone else's expectations.
Mirror is an AI-powered career and major advisor that builds a real psychological profile of each student through natural conversation, not generic quizzes. It asks the questions a great mentor would ask, remembers the answers, and delivers tailored recommendations grounded in real data.
Most students don't choose the wrong major because they lack information, but because they lack self-knowledge, or because they're quietly performing someone else's expectations.
Mirror is an AI-powered career and major advisor that builds a real psychological profile of each student through natural conversation, not generic quizzes. It asks the questions a great mentor would ask, remembers the answers, and delivers tailored recommendations grounded in real data.
My Role
My Role
Product Designer: I owned this end to end. The problem, the concept, the research, the visual direction, and every decision about how the product would work.
AI Interaction Designer: I designed how the user talks to the AI. The persona, the conversation flow, the guardrails, and the system prompt behind all of it.
Design Engineer: I built and refined the interface and backend in code, primarily in Cursor and Claude Code.
Product Designer: I owned this end to end. The problem, the concept, the research, the visual direction, and every decision about how the product would work.
AI Interaction Designer: I designed how the user talks to the AI. The persona, the conversation flow, the guardrails, and the system prompt behind all of it.
Design Engineer: I built and refined the interface and backend in code, primarily in Cursor and Claude Code.
Outcomes
Outcomes
90% Accuracy
90% Accuracy
Nearly all tested users who already had a chosen major reported that Mirror accurately gave them similar majors/careers to what they chose.
Nearly all tested users who already had a chosen major reported that Mirror accurately gave them similar majors/careers to what they chose.
Discovery
Discovery
A problem of identity, not information
A problem of identity, not information
Who did I want to help?
Who did I want to help?
I wanted to help students who feel lost, who changed their major, or who got pushed into a path that never felt like theirs. My target users were 17 to 20, of any background.
When I researched the existing career and major quizzes, they all had the same flaw. They ask you to self-select interests from a list, then hand you a generic result. The students who need help most are the ones who don't yet know what they like.
I wanted to help students who feel lost, who changed their major, or who got pushed into a path that never felt like theirs. My target users were 17 to 20, of any background.
When I researched the existing career and major quizzes, they all had the same flaw. They ask you to self-select interests from a list, then hand you a generic result. The students who need help most are the ones who don't yet know what they like.
Listening to real students
I interviewed six students who had either switched majors or faced outside pressure around their choice.
The same themes came up again and again. Family expectations shaped their decisions. They didn't feel like they knew themselves well yet.
Almost none of them had used a career quiz, because the ones that exist felt too shallow to bother with..
I interviewed six students who had either switched majors or faced outside pressure around their choice.
The same themes came up again and again. Family expectations shaped their decisions. They didn't feel like they knew themselves well yet.
Almost none of them had used a career quiz, because the ones that exist felt too shallow to bother with..
Turning interviews into direction
Turning interviews into direction
Two clear signals shaped Mirror. Students wanted something conversational, not a button-clicking quiz. And they wanted to watch it understand them, so their answers didn't disappear into nothing.
I built two personas from these insights and used them as a gut-check for every design decision that followed.
Two clear signals shaped Mirror. Students wanted something conversational, not a button-clicking quiz. And they wanted to watch it understand them, so their answers didn't disappear into nothing.
I built two personas from these insights and used them as a gut-check for every design decision that followed.
Design
Design
Working backwards on purpose
Working backwards on purpose
Refining in Cursor
Refining in Cursor
I knew exactly the look and feel of the interface that I wanted, including fonts, colors, even layout before ever touching any tool.
I knew exactly the look and feel of the interface that I wanted, including fonts, colors, even layout before ever touching any tool.
Taking a new approach
Taking a new approach
I'd always worked wireframes, then Figma, then dev. This time I flipped it: I wrote a concept, turned it into a PRD, and used Claude Code to build the whole app first, then refined subtractively. I knew the tradeoffs of skipping wireframes, but I was confident I could refine from an existing design.
It was a test of my UX brain. Could I look at a finished, imperfect interface, pinpoint what was wrong, and iterate? I split the PRD into nine phases, and created a markdown file explaining how to implement each phase so the build checked in at every step instead of vanishing into one giant prompt.
I'd always worked wireframes, then Figma, then dev. This time I flipped it: I wrote a concept, turned it into a PRD, and used Claude Code to build the whole app first, then refined subtractively. I knew the tradeoffs of skipping wireframes, but I was confident I could refine from an existing design.
It was a test of my UX brain. Could I look at a finished, imperfect interface, pinpoint what was wrong, and iterate? I split the PRD into nine phases, and created a markdown file explaining how to implement each phase so the build checked in at every step instead of vanishing into one giant prompt.
The initial build
The initial build
Once Claude Code was done, the inital build was at 50 percent of what I envisioned this whole product to be. I moved the whole project to Cursor where I could fine tune and make more specific edits.
Cursor also gave me better control over the backend and also was easier to eventually publish on Vercel.
Once Claude Code was done, the inital build was at 50 percent of what I envisioned this whole product to be. I moved the whole project to Cursor where I could fine tune and make more specific edits.
Cursor also gave me better control over the backend and also was easier to eventually publish on Vercel.
Fixing the results page
Fixing the results page
The first results page was an information overload. I knew my users had shorter attention spans, and I knew it needed to be overhauled to displayed only the things that were important to the user.
The first results page was an information overload. I knew my users had shorter attention spans, and I knew it needed to be overhauled to displayed only the things that were important to the user.
Conversation
Conversation
Crafting a intuitive AI experience
Crafting a intuitive AI experience
Designing the conversation
Designing the conversation
This was the heart of the project, and the first time I had ever designed an interaction with an AI.
I designed Mirror's persona to feel like a warm, grounded mentor. The conversation moves through three phases, but the user never sees a label or feels a transition. From their side, it just feels like a really good conversation.
This was the heart of the project, and the first time I had ever designed an interaction with an AI.
I designed Mirror's persona to feel like a warm, grounded mentor. The conversation moves through three phases, but the user never sees a label or feels a transition. From their side, it just feels like a really good conversation.
Tuning the AI itself
Tuning the AI itself
Getting the conversation right took real iteration. Early on, Mirror asked too many follow-ups and lost track of where the user was. I also caught it leaking its internal reasoning into the visible chat, which I fixed in the system prompt.
Model choice mattered more than I expected. I tested across models and stuck with Opus, since its reasoning followed the system prompt far more reliably than the others.
Getting the conversation right took real iteration. Early on, Mirror asked too many follow-ups and lost track of where the user was. I also caught it leaking its internal reasoning into the visible chat, which I fixed in the system prompt.
Model choice mattered more than I expected. I tested across models and stuck with Opus, since its reasoning followed the system prompt far more reliably than the others.
Questions that reveal, not interrogate
Questions that reveal, not interrogate
Instead of asking "do you like math," Mirror asks scenario-based questions that surface personality without the user realizing they're being assessed. Things like walking through your ideal free Saturday, or how you'd handle two teammates who disagree.
Instead of asking "do you like math," Mirror asks scenario-based questions that surface personality without the user realizing they're being assessed. Things like walking through your ideal free Saturday, or how you'd handle two teammates who disagree.
Watching yourself take shape
Watching yourself take shape
As the conversation goes on, a profile sidebar fills in live: values, work style, interests, pressure flags. It gives users a sense of progress without a progress bar. You're not on question 12 of 20. You're watching a portrait of yourself form.
As the conversation goes on, a profile sidebar fills in live: values, work style, interests, pressure flags. It gives users a sense of progress without a progress bar. You're not on question 12 of 20. You're watching a portrait of yourself form.
Keeping it in scope
Keeping it in scope
I scope-locked Mirror to career and major advising only. When users go off-topic, it redirects warmly but firmly, like a focused mentor who stays on task because they care, not a system hitting a wall.
I scope-locked Mirror to career and major advising only. When users go off-topic, it redirects warmly but firmly, like a focused mentor who stays on task because they care, not a system hitting a wall.
Delivery
Delivery
The final build
The final build
The landing page and chat interface
The landing page and chat interface
The conversation
The conversation
The conversation takes around 15 minutes. Mirror asks questions and adjusts as you go to help itself figure out ideal career paths.
The conversation takes around 15 minutes. Mirror asks questions and adjusts as you go to help itself figure out ideal career paths.
The results page
The results page
When Mirror is ready to build your results, the interface will transform into four potential career paths with more information on each card.
When Mirror is ready to build your results, the interface will transform into four potential career paths with more information on each card.
Export as a PDF
Export as a PDF
Users can export their results as a PDF to give to parents, or school counselors to help figure out the right fit for their career.
Users can export their results as a PDF to give to parents, or school counselors to help figure out the right fit for their career.
Reflection
Reflection
My Key Learnings
My Key Learnings
What did I learn?
What did I learn?
Designing for AI is designing a personality.
Designing for AI is designing a personality.
Designing for AI is designing a personality.
This was my first time shaping an AI interaction, and most of the work lived in the system prompt, not the UI. Tone, pacing, when to push, and when to back off were all design decisions.
A finished build is just the starting point.
This was my first time shaping an AI interaction, and most of the work lived in the system prompt, not the UI. Tone, pacing, when to push, and when to back off were all design decisions.
A finished build is just the starting point.
A finished build is just the starting point.
A finished build is just the starting point.
Building the whole app first and refining backwards felt strange at first, but it sharpened how I spot experience problems in something that already exists. The real product came from the iteration, not the first generation.
Building the whole app first and refining backwards felt strange at first, but it sharpened how I spot experience problems in something that already exists. The real product came from the iteration, not the first generation.
Thanks for Reading!
Thanks for Reading!
Check more of my projects, or reach out if you want to learn more about them!
Check more of my projects, or reach out if you want to learn more about them!
© 2025 morganerdman.com