project Highlights
From idea to passed exam in 7 days
Adaptive engine: mastery-weighted questions, exam-readiness estimate, per-topic mastery map
78-term search alias index
Personal brand & design system running in production (WCAG AA, semantic tokens)
7 documented releases with in-product release notes
Built for 1, adopted by the class & everyone passed
project Deliverables
Live platform (seven study modes), content architecture & question bank, design-system implementation, public product release notes.
Context
Halfway through my CMD programme at HAN, a theory exam was coming up: Human-Centered Design, consisting of four subjects. I know that passive rereading doesn't work for me, I learn more by active retrieval. For an earlier exam I'd built myself a single-file practice quiz. This time I wondered if I could do better and accidentally build a full learning environment.
Shipping a core version, based on the course curriculum
Shipping the core first
Working in Claude Design, every prompt had to be fully self-contained, the build model has zero context from my planning sessions, so each brief carried the story, intent, architecture and changelog. My previously build my framework (SPINE) and it ran a pre-flight on every big version update.
v2 shipped the first working product: seven study modes (including video's, reading, quizes (open or multiple choice questions) summaries or flash cards), all content derived from the course material (4 subjects), all progress stored locally so to avoid waisting time building user accounts. Deployed as a static SPA on Netlify, based on my personal brand and design system.
Making it adaptive
v3 was the pivot from content delivery to modelling the learner. The questions adapt to your answers, to track your progress and challenge your weaknesses. An exam-readiness estimate with advice on where to spend your time, and a mastery scale per topic. But then…
Adding a model that analyses readiness based on the answers and advices where to focus on next.
Then the class shows up… and breaks the search
Once classmates started using it, and I added an in-app feedback channel and a search functionality. I learned that building a search is not as straight forward as you'd think.
From feedback and testing I discovered that searching for things like "five users", "HCD", "informed consent", terms that weren't literally in the content, but highly related to the content, did not lead to search results.
The fix became v5.2: a 78-term alias index covering synonyms, English terms, abbreviations and spelling variants, with type-ahead that resolves "informed" to Informed consent. A closed research loop inside a side project. It's still not a perfect search, but for the scope, it did the job.
I have new respect for search engines, building one a league on it's own.
What worked?
Faster, targeted studying
Classmates adopting a tool built for one
What didn't work?
Search only matched literal course wording
Real study vocabulary (English terms, abbreviations) came up empty
Polish and… showtime!
Later releases (Release notes) added a lecture-video watchlist, Enter-key flow-through with input-field guarding, and an expansion of the theory itself (e.g. Norman's interaction principles, visual design principles, ISO 9241-210), every new addition also created automatically the corresponding flashcards and search. That propagation was the architecture proving itself.
Results
Everyone in the class passed; I scored an 8,7
Seven days, seven releases, all documented
The stack made my €20/month build subscription redundant (now cancelled)
What I learned
Good AI starts with good information architecture. Organising the course material in Notion before prompting meant the build model received a clear information architecture instead of having to invent one. This also helped with analysis for the search index, later on.
A fixed deadline is a better design constraint than perfectionism. A fixed exam date forced me to focus on what mattered most. Building in small iterations between assignments kept the scope realistic and prevented endless polishing.
My role shifted from designing to directing. Instead of designing myself, I focused on defining the problem, giving direction, reviewing outcomes and deciding what deserved another iteration. Having a brand and design system (however small) made all the difference.
Release notes are part of my AI-assisted workflow. For this project, I shared them with my classmates to keep everyone informed of each update.
Final version with the full curriculum, videos, search, and adaptive learning. Built alongside my studies by directing Claude Design in parallel.
Different methods of learning, everyone has a different preference.