
Course Companion
My side experiment became a mission, and I led its design and build. One shared file anyone could build in, a live test, a fast turn when money did not follow.
- Year
- 2026
- Track
- built
- Category
- case-study
Tools
- Claude Design
- Figma
- Next.js
- Cursor
I rebuilt a course page so what the AI makes lives on it, for anyone taking the course. It went live on 22 May as an A/B test against the old page, with the first version, and grew in steps to 10,127 courses. Reach into the AI tools went from about 8% to about 31%. Fewer students reached the documents, where the company earns, so by 20 August we had turned to documents first. The picture above is my rebuild prototype on sample data. The page that went live is further down, and the rebuild reached students in pieces after it.
It started as my side experiment
I always thought our distribution model was unsustainable. So in early 2026 I built a different course page on my own time, and on 25 February I pitched it to the management team.
The company had built an AI-native product. It took what a student typed and the documents they gave it, made practice questions, summaries and mock exams out of them, and saved those to the student's private project. In its beta program, most users left and didn't come back, and they were confused by how the private projects worked. That was being treated as a marketing problem. I thought it was structural. The AI lived in the wrong place.
In the pitch I called it a vending machine. A student walked up with their own material, got a summary or a mock exam out, and left. It did the job, and it was lonely. Nobody else taking the course ever saw what came out, and nothing it made stayed behind for the next student. It all went into one student's private project, scattered where nobody else could reach it, and soon its maker forgot it too. We wanted students to find what was relevant, not make it from scratch every time. Students do not go looking for an AI tool. They go to their course page, because that is where their material and their classmates are.

The AI product's front door, as it looks in October 2026: bring your material, pick what comes out.
The course page itself was largely neglected. It had seen lots of small changes, but the company's money ran through it, so nobody risked a serious experiment on it.
We'd forgotten how the company grew: students uploaded their documents, we put them on public pages, those pages brought in new students, and some of them uploaded too. The AI sat outside that loop, so it brought nobody new. I wanted to put it back in.
So the pitch was one change: what the AI makes lives on the course page, not in a personal workspace. A student makes a mock exam from a lecture note, and it sits on that course's page, for anyone taking the course, signed in or not, this semester or next. A shared link carries it into the class group chat, so the page does not have to wait for search to send people. The pitch proposed moving the paywall with it, to the moment a student fails a practice question, because that is the most motivated user in the product. Reading and making stay free. Feedback is the paid part.
A month later there was a mission and a small team, and I was the designer on it. It was the company's first mission, a pilot, and it changed how the company worked.
The course page students had

The reframe, from the pitch
The root of the business
Shared documents fed the platform back
This is the system the company was built on, and it worked.
The course page
Years of uploads
- A student shares a document
- It goes public, on the course page
- Other students find it on the platform
- Some share their own, and the course page gets richer
Then round again
The AI product
The AI was a vending machine
It did the job, and it was lonely.
- A student
- Their own material, or a prompt
- The AI tool
- A quiz, summary or mock exam
- Their private project
- Scattered in private projects, forgotten by its maker
What I proposed
What the AI makes goes back into the loop
It lives on the course page, not in a private project.
The course page
For anyone taking the course, this semester or next
- A student makes a mock exam with the AI
- It goes public, on the course page
- Other students find it on the platform
- Some make their own, and the course page gets richer
Then round again
A documentA mock exam
The demo it ended on · Prototype · sample data

Three of the four students reached for the quizzes before the documents
In April and May we showed the first version to four students. Our main hypothesis was that students do not care whether something is a PDF, only whether it is useful, so what the AI makes from their course material would be worth at least as much to them as an upload. It only ran inside the company network, so we drove while they talked. Three of the four reached for the quizzes and mock exams before the documents.
The exam countdown was the first thing every one of them read, and asked what the page was, they described a place to prepare for the exam.
Nobody searched the page for what was already there, but they took what we put in front of them.
It went live on 22 May, and the test grew in steps
It was eight weeks after the mission started, and we had misjudged the traffic.

The first version, as it ran on 1 June, ten days after launch. The page the four students had tested.
None live
No students
Inside the company, behind a switch, so the old page could not break.
2 courses
1 in 10 visitors
Live on the date we set, with a switch to pull it.
161 courses
Half of visitors
More students for a clear read, on courses with confirmed exam dates.
978 courses
Half of visitors
Only courses with exam dates. The rest waited for a design.
10,127 courses
Half of visitors
The top US courses by traffic, still chasing a clear read.
Today carried the traffic, so on 23 June we locked the rebuild around it
We opened the page on Today because students struggled to find what mattered on their course page. By June we had their feedback and a month of the live test: the Today page carried most of the traffic to the AI tools, the sidebar almost none. On 23 June the PM and I locked the rebuild: a modular page, tabs instead of a sidebar, one main column.
In the rebuild, Today opens on a question, with no start button. The daily review takes three topics in about ten minutes, weakest first, and under the card is the plan, the course laid out toward the exam.
Before: Sidebar navigation
After: Mobile-style tabs, everywhere
Before: Browse by material or topic
After: A prioritized spine: Today, Tomorrow, milestone
Before: A start practice click
After: A question the moment you land
Before: Static progress bars
After: One coverage ring, how much you know
Before: Sessions disappear
After: Topics persist as navigable artifacts
Before: Passive materials list
After: Community pool and a personal study layer
Six students read the page the way April's four did, and four things changed
Writing April's round up, I found the generated notes had stopped before the prototype came on screen, so I rebuilt the findings from the full transcripts. In the second round, the students drove on their own screens.

The page we tested with the six students in July, on a made-up course.
In July I led six sessions, three of them from the first round, and product ops and the PM joined some. The countdown read first for all six, and all six read the page as the place to prepare for the exam.
Come back tomorrow, which I meant as spacing, read as scheduling and gating.
Four things changed.
The hub landed. Students read the page as the one place for the exam.
A student: “Everything is organized really nicely. Like, I don’t have to do any of the organizing, and it’s just in front of me.”
What changed: The Today page got more lanes, documents and quizzes beside the questions.20 July
Trust was the gate. Every student wanted a reason to believe the AI prep: a source, the lecturer’s view, the topics it covered, or other students’ ratings.
A student: “I could only rely on the theoretical part of the exam. For the calculation part, I couldn’t actually trust it.”
What changed: Claims got a source or went.From 28 July
After a wrong answer, students wanted a score and a way to strengthen the topic, not just more questions.
A student: “If I click this, it would go back to the lecture that it’s associated with, and then maybe flashcards.”
What changed: The result screen kept its score, and after a wrong answer the page showed documents to read, not another question.21 July, then 4 August
Endless practice split them, and the ones it lost wanted an end and a score.
A student: “I would like to see the score actually, how good I’m performing. It kind of gives me more motivation.”
What changed: I put cram mode forward in its place, and the concern went to the practice team.20 July
One more change came from a designer on another product: we fitted the page to her team's practice flow instead of my standalone prototype, and I changed my build to fit.
The countdown held, but on the calm page one student said his midterm did not feel important, and the exam-forward version was what he was looking for.
So cram mode went forward. It changes the whole Today: one card with the countdown in place of the exam tiles, Highest-impact first in place of the daily review, every weak topic queued and no Come back tomorrow, Rapid recall with one fact a topic, and only what is before the exam under it.

Prototype · sample data
Cram mode, the whole Today, in the late-July build. Scroll inside it.
Go deeper
July's sessions were thirty-minute video calls: the student shares their screen and drives, five minutes on how they study now before any prototype, then quiet exploring, then questions only about what they did not reach. The rule: test the page, not the person. Confusion is data. Do not rescue, defend or sell.
The read-out's own caveats: made-up data, stated intent only, medium confidence, no return behaviour observed. Two of the six also worked at the company.
Today's review




Prototype · sample data
The loop as students saw it in the July sessions, before the research. Come back tomorrow is the mechanic they read as scheduling.
Where to next, after the last question, as it stood in the research build. Few reached it. The hyped card is the infinite practice that split them.
Where to next
5 states

Where to next? Pick whatever fits your energy right now.
Biological Psychology explained. Explainer, 8 min.
Infinite practice. Adaptive questions that never run out.
Biological Psychology, 12-min audio recap. Podcast, 12 min.
Biological Psychology, full reading notes. Lecture notes, PSYC 101.
Prototype · sample data
The two probes written for Come back tomorrow.

Probe 4: what does Come back tomorrow do, and does it feel like help? 
Probe 5: why hold these for tomorrow, and do you trust it to decide?
After July, a new student started with a readiness check, weakest first
A new student first gets a readiness check across every topic, and its result ranks them weakest first. The daily review starts from the weakest, and the score after each sets the push: Strengthen this topic when it's low, Sharpen when it's close, Go deeper past ninety.

Every topic gets a score, so the plan can start from the weakest.
Prototype · sample data
Under Today's card is the plan, the course laid out toward the exam.
Today
Three topics, weakest first, in about ten minutes.
Tomorrow
The same topics come back before you forget them.
Recently added
What classmates uploaded or made with the AI this week, with who added each one.
Before the exam
A readiness check that scores every topic. It counts down once we hold the date.

Prototype · sample data
The plan as I designed it for the best case, a course we hold everything for and a student with history: one column from today's review down to the exam. Scroll inside it.
We built the options instead of arguing about them, then chose by a rule
The team's model for not knowing enough about a course was one dial, four levels, from we know nothing to we know everything. It held until it had to describe a well-covered course seen by someone on their first visit: one dial has to either invent a readiness score from no history or hide dates we actually hold.
So at the end of July I split it: what we know about the course on one axis, what we know about the student on the other, sixteen states. Down a row is a content problem, and nothing a student practises moves it. Across a column is an engagement problem, and only the student moves it. Four things change from state to state: the daily card, the plan, the topic radar and the one ask we may make.
We never drew all sixteen: it was a map, and the prototype renders each state from the rules. Then I built a check that grades the page against the map, written before the page it grades, so it could not be tuned to flatter the work. Then I fixed its own false greens.
Confidence matrix
16 states
- L1 we hold no topics
- L2 everything between
- L3 a confirmed course and a student with history
- The setup
- We hold no topics for this course, and this is a student who just arrived and has done nothing yet.
- So we show
- the Must-know deck. It only needs documents, and that is all we have.
- We ask
- for the exam date, since we have none.
- Daily cardMust-know, key facts from the most-studied materialmatches
- SpineLocked preview, 3 dashed rowsmatches
- RadarHiddenmatches
- The askExam date, derived from an existing mock’s namematches
Each row is the plan, written before the page. Matches means that kind of card, plan, readout or ask is on screen.


Prototype · sample data
The confidence matrix, my prototype on sample data, open on a course we hold nothing for, seen by a student who just arrived. It shows what the material allows, asks for one date and says what the plan needs. C3 × S0, the page at the top of this case, is the cell that forced the split.
Under the states are rules for what the page may say when it does not know. A field shows only if we can produce a value and that value drives something. An estimate carries a mark you can press, to see the document it came from and correct it. With no exam date, we ask blind, because showing our guess first changes the answer. We ask only about something you are looking at, never a cold audit of a list. A student with no history has no readiness score, and the readout says so rather than inventing a number. The split, the rules and the picks were mine.
The same move, smaller, came for the exam date. Most courses had no exam date to show, and the card we handed to engineering said dates are not in yet: true, and nothing for a student to do. So we ran a session on it, the team in the room. The premise was that a student is always walking toward something, so the slot is never empty, only empty of exams. The options came as a ladder, each rung needing less data than the one above, and the rule was to fall to the next rung only when the one above is empty.
After review the ladder got shorter: school calendars turned out to be rare, so that rung fell, and the ask shrank to one date. Engineering got the top rung, the syllabus's real dates. Six other designs for the strip stayed one step down, in the panel where you flip between them.
Go deeper
The one dial, as the team had it before the split.

The one dial the team started with: four levels, from we know nothing to we know everything.
Prototype · sample data
On 21 July we decided readiness belongs in the header next to the exam dates, and left the shape open. I built all nine ways to combine them behind one switch. They were being tested with another team and with students, with no pick made. That is the honest cost of building first: a switch puts off a decision, it does not make one.
The course header
9 states

The current header.
One bar, an 80% “exam-ready” flag, the date beside it. Lowest cognitive load.
Two stacked bars: time-to-exam vs your readiness. Closest to my sketch.
Blue fill (you) on a gray time-elapsed ghost, one bar. Its line comes closest to a pace judgment, where the others state a gap and the days left.
Bounded height, checkered finish, one accent.
A mini timeline, the climb from 52% to 80% by Apr 8. Most expressive. In this build two labels on the chart lost their numbers.
Readiness over the last 8 weeks, trending toward exam-ready. Backward-looking momentum where the other timeline styles are a forward climb.
A calm take on a study-consistency streak, in the design system's own parts (not a Strava clone).
The exam-strip tile we handed to engineering, with readiness merged in, and the cheapest variant to build and test.
Prototype · sample data
Under every treatment the rule is the same sentence: the nearest true thing we can put in the slot.
The exam strip, with no dates
7 states

What we had handed to engineering: one honest “dates aren’t in yet” card. True, and nothing to walk toward.
The exam strip one confidence step down: same cards, the status-warning ramp carrying the confidence. Set it or leave it.
One 62px track: confirmed coursework as a chip, the roughly-known exams as pressable pills sitting on the rail.
Today’s strip shape with three states: pinned, roughly known (Set day), and set by you, which this frame does not show. Nothing new to learn.
Collapsed by default. One row: the next thing up with its countdown, the rest of the term as a quiet line, and “3 more this term” unfolds the real month grids (not shown), where a week we only know roughly is a band across that week.
The same months, no day grids, monochrome: each item carries its kind’s glyph, a “?” marks a date we are guessing, and every box hugs its own content. One item is ~58px, a second grows it. The + on a month adds one; a month holding nothing is a thin +.
The brainstorm’s converged band, fully chronological: the next three term items by time (exam, quiz, essay, whatever comes first) as tab pills on one centered line, with a ≈ countdown chip on the nearest, the only blue. A readiness readout while collapsed, and everything further out behind +N more, where each row expands in place: covers, format, notes, when. Scraped from past terms’ syllabi, so every date reads ≈ and carries a Set-the-day ask.
Prototype · sample data
In front of the Documents tab there were six orders, the old library reachable from every one of them as the last line under the results, never the first thing.
In front of the Documents tab
6 states

The legacy library grid, untouched, in no particular order. The control. In this build it shows four documents.
A find-in-this-course box over type facets, the one thing the flat grid cannot do. Best when the student already knows what they want.
One ranked rail per kind, best-rated first. “I want a practice exam” becomes a glance instead of a scan.
Documents mapped onto the course’s topics, so “which doc covers X” is answerable. And it says out loud that the mapping is our guess, filing the rest under the whole course.
Six documents, each with the reason it earned a place: opens, ratings, this week. The library is a scroll; this is an answer. Browse-all is one click behind it.
The same grid with an honest default sort and facets. Nothing new to learn, and the cheapest of the five to ship.
Prototype · sample data
Reach went up, the money did not follow, and we turned fast
Reach into the AI tools went from about 8% to about 31% among students who saw the new page, and the share back in the first week went up about ten points. The figures are relative, because the absolute ones are the company's.
The cost was on the other side. Fewer students reached the documents, and the documents are where the company earns. That was April's finding at scale: three of the four had gone to the quizzes before the documents. Purchases once a student reached a document stayed the same, so it was a distribution loss, not a demand loss. The AI side did not pay for itself. The paywall never moved.
On 4 August the PM proposed putting documents back wherever there was no AI content yet. I objected to that being our limit: with documents as the ceiling, we had rebuilt the page we set out to replace. The PM kept it as a trade-off, and I built it the way July's sessions had pointed: documents inside a practice set only after a wrong answer, and sharing at the end. By 20 August, with capacity down to one engineer, it was documents first, with a page that falls back one piece at a time, so a missing part never empties it.
What this course holds
3 states
Practice questions exist, so the daily review cards lead. Today's page.
No questions yet, so Today's review is gone and the spine opens at the exam milestone with the AI artifacts and documents we do have.

Not even artifacts, so that first section becomes Top documents and carries document tiles only.
Prototype · sample data
Here are six of the ways the prototype put documents under a question a student got wrong.
Under a wrong answer
6 states
One line under the explanation: how many documents in this course cover it, and a link to them.
One tile for the session with a count, not the same tile under every question.
When the document is one the student uploaded, the tile says so.
The course's own words under the explanation, then the document they came from as one tile.

Only after a wrong answer: two pages from the course that cover it.
On the result screen: the pages for the questions missed come first, and the ones saved sit behind one line.
Prototype · sample data
Go deeper
When the questions run out, the page has one thing worth sending: the set a student just did, and their score on it. So the ask to share comes once, under the score, and never beside another ask.
The share loop





Prototype · sample data
The asks came out of research I did in August, on how Gen Z gets let in to clubs, festivals and parties. Its one thing not to build was share-to-unlock, and none of it was tested with students.

The first three of its nine rituals. The stamp, the bands and the stickers in the window below came from them.
Seven of the fifteen asks move. Press one to see how.
Four of us worked from one file, and anyone could change it
There were four of us. I led the design and built the prototype. Our acting PM was also the Head of Design, so we designed together and gave each other feedback the whole way through. A frontend engineer transformed what I built into production code. Product ops and growth were one person, shared with other teams: she ran the numbers, sat in on some of the interviews and brought the ideas for getting the page in front of more students.
At kickoff we agreed to skip conventional requirements documents and keep them as live artifacts instead. When the team needed to agree on something, we pointed at the prototype. I had rebuilt most of the company's design system inside the file the prototype lived in, and the whole team worked from it.
I coded the rebuild's first states with AI and kept shipping production code through the summer, with the engineer rebuilding and hardening on a shared branch. The proof of concept went in as seventeen commits, all mine.
We did not have design handoffs: the source of truth was always up to date. That said, on 2 July our engineer told us he could not match the prototype in the browser: the colours and sizes were not named the way the codebase names them. So I built one link: the prototype, every component with the px and hex the browser could match, and the product context, forty-two pages. Its cover says handoff, but it was that source of truth.
Beside the file sat the team's second brain, a shared set of notes. Scripts on a schedule filed the meeting notes and posted a Friday digest, so everybody had the latest decisions without asking. How it works is its own entry: the self-maintaining vault.
We saw the new page as the opposite of the old one: a place to try bold things, ours and the other teams'.
The course page
The hub, where a student starts to study
- Documents students uploaded
- Quizzes, flashcards and mock exams the AI made in this course
From the practice team
Their topic list
Our page built on it and did no extraction of its own. Our courses got their topics first.
22 June and 15 July
From the practice team
Subtopic data
We turned it into review questions for each topic.
To the practice team
Today's check
It led topic by topic into their endless practice, with a mock exam to close.
From 29 June
With the practice team
The exam readiness score
We fed it to them and fetched it from them.
With the course-data team
Our need for real exam dates
By August the confidence model and the exam-date layer sat on their side.
From 15 July

From 27 July I kept a tracker of where the teams met: handoffs, blockers and work done twice. A job I scheduled rebuilt it every Monday from the teams' boards, channels and docs, for anyone to read, and corrected it in the open. The top of it on 3 August, with the other teams' names replaced.
This is how I work, on this project and since. I do not sit in front of AI tools and wait for designs to come out. Our tools had standing instructions to go to the second brain first, and only then answer. I go back and forth: hand sketches, Figma, variants and states in Claude Design, and code when a state has to be real. I share the result as a working prototype with teammates and other designers, then put it in front of students. I threw out at least half of what the tools made, because it was too complicated to look at or outside our design pattern.
Go deeper
Three more pages of the link.

The topic card, collapsed, in each of its states. 
Its answers in every state, and the words each level of coverage earns. 
The same card's measurements, named so the browser can match them. 
How Today's review picks what to show in the time a student has. The weights are covered. 
Six of the eight empty and edge states, with when each shows and what it says.
Anyone on the team could add an idea to the prototype
I wanted the file to be something the team built in, not something they waited on. So I wrote prompt templates: a form that writes a ready prompt for anyone to add an idea to the shared prototype, with the house rules already in it. Anything new is switched off until we turn it on, so the prototype stays as it is for everyone in the meantime. The house rules are the design system's colours and type, nothing smaller than 14px, and the project's notes updated in the same turn. The PM built header variants that way, and product ops proposed different layouts that the PM and I later took in. We spoke one design language, and anyone could make options in it.
The same went for research. I built an interview companion so anybody on the team could run a session with a student: the script one step at a time on a thirty-minute clock, a consent line to read out and then wait for a clear yes, neutral prompts that do not lead, and a list of things not to say.
Go deeper
The form writes three kinds of prompt:
- A new feature, behind its own switch. The default.
- A new version of the prototype, copied from the current one. The old page stays standing beside it, so the two can sit side by side in an interview.
- A new look for something that already has a switch, opt in, so a style choice never becomes a new switch.
Before each session there were two things to set up.

A session set up in the prototype's own panel, on sample data. The link opens the clean student view with exactly these settings, one prototype per student. 
The companion before a session: who, which course, and a checklist for the build and the recording.
The other steps hold the things not to do (no "do you like it", no feature names before they use them, no rescuing at the first pause), the probes, and a button that copies the notes out.
What I carry forward
Most of this work was deciding what a screen may say when it does not know enough, and the honest answer was always more work than the dishonest one: ask instead of guessing, show a dash instead of a number, build a page that falls back one piece at a time instead of assuming the data arrived.
The other half was how we worked. One file everyone could change, a second brain that kept everyone current, prompts so anyone could build, and a prototype in front of teammates, other designers and students before anything was written down. I would set that up again on day one.
Go deeper
From the pitch to where the record stops, to scale.
- I pitched a new course page to the management team and ended on a demo.
- It became a mission with a small team, and we kept the requirements as live artifacts instead of conventional documents.
- The team's first running version shipped inside the company, on the date we had committed to at kickoff.
- – We showed the first version to four students from our research panel.
- It went live as an A/B test against the old page.
- The PM and I locked the rebuild.
- –I led the second round of interviews, with product ops and the PM joining some, putting the loop in front of students before it was built.
- –I split confidence into two axes and mapped the sixteen states.
- – I built the check that grades the page, then fixed its own false greens.
- Where the record stops, mid-sprint.
More case studies

Workwize Onboarding Case
Workwize's design case for a Senior Product Designer role: a delivery date means nothing until it is judged against the day the new hire starts.
case study · 2026 · read it
Soothe
A native iOS proof of concept that pats you until your stop is near, then wakes you kindly. A 2026 rebuild of my 2020 case study. Every ride is simulated.
case study · 2026 · wip · read it
AI Flashcards
A short-term vision for the flashcards in Studocu AI: a deck made from a student's own sources, one rating per card, and a deck that says what to study next.
case study · 2025 · read it




