A case study about learning, not streaksProduct design · 13 weeks · 2026

A 140 day streak,and I still couldn'torder a coffee.

GermanlyA German learning product for international students and workers in Germany

Most language apps are designed to bring you back tomorrow. Very few are designed to make you able to speak. This is the story of why people fail with apps built for retention, and what I changed when I designed Germanly.

120+learners today
3,000+impressions on day one
mygermanly.com
In numbersAfter 13 weeks and one redesign

I designed and built Germanly alone, from the first interview to the redesign. It now teaches 120+ learners, and a 20 person pilot study, now under review at CHI, showed learners felt more confident and treated it as something they own.

CHAPTER 01The streak

I kept the streak. I still couldn't speak.

I moved to Siegen for my HCI Master's. Flat viewings, the Ausländerbehörde, the doctor, the bakery: every one assumed I could hold a basic conversation. I did what everyone around me did. I opened Duolingo, kept the streak going for months, and still froze at the bakery counter.

My setup looked like everyone's I later spoke to: fourteen browser tabs of YouTube lessons and grammar blogs, three half finished apps each good at one thing, and a folder of PDFs I never opened twice.

The apps were not broken. They did exactly what they were designed to do: bring me back every day. Nobody had designed them to get me speaking.

CHAPTER 02Why people fail

Apps built for retention measure whether you came back. Not whether you can speak.

A retention loop is short and satisfying: a reminder, a few taps, a reward, a streak. It is excellent at bringing people back. But recognising a word on a tile is not the same as recalling it at a counter, and nothing in the loop checks the difference.

The retention loop

Optimised for coming back

Each step is easy, so the loop is fast. The learner rarely produces a sentence, rarely meets an article under pressure, and never learns whether they are ready for an exam.

The learning loop

Optimised for what sticks

Each step asks for a little more effort: recall instead of recognition, a sentence instead of a tile, a readiness score instead of a flame. It is slower, and it is what the exam and real life measure.

CHAPTER 03The people

Fourteen people, one deadline: B1.

I wanted to know whether this was my problem or everyone's. I spoke to international students, newcomer workers and professionals about to move. Almost all of them had a date in the calendar: a visa appointment, an admission, a residence permit.

In the room · workshops and interviews

How I ran the interviews

    What kept coming up

    Meet the 20 learners from the pilot study
    IDAgeFromBackground
    CHAPTER 04What Germanly changes

    Keep what makes apps pleasant. Change what they measure.

    I did not want to build a harder app. I wanted one that is just as easy to open, but where every tap moves you toward speaking and toward the exam. Five shifts came out of the research.

    CHAPTER 05Thirteen weeks

    From a Notion page to 120 learners.

    Three phases, one designer. I researched before I built, launched as soon as the path worked, and let what learners did after launch decide the redesign.

    CHAPTER 06The blueprint

    Before pixels: the methods, the map and the paths.

    These are the UX foundations under every screen: which methods I used and why, how the product is organised, and the paths people actually walk through it.

    The UX toolkit, and why I used each method

    Information architecture

    The navigation had grown to fourteen destinations, one for every feature. I regrouped it around what people do: come back, follow the path, look something up, practise.

      User flows

      Four paths carry almost every session. Pick one to walk through it.

      CHAPTER 07Design decisions

      Nine design decisions, each traced back to research.

      Each decision below shows the product in action, then why I made it, the brand rules it follows, how AI helped, what the API taught me, the design system step it created, and how it compared with other apps.

      CHAPTER 08Style guide and system

      One style guide keeps every screen consistent.

      After launch I measured every screen. One page had 431 controls and no primary action; the lesson had 19 and exactly one. The fix was not another visual pass. It was a system: six tokens, grammar colours with a fixed meaning, two typefaces, eight rules and eight components.

      Colour

      Dark first, because learners study late and on phones. Colour is rationed: one tinted card per screen, activity colours only for progress, blue only for links and focus.

      Typography

      Satoshi for display, Switzer for reading. Tracking tightens as size grows, which is what makes the large numbers feel calm instead of loud.

      Spacing

      Radius

      Components and their states

      Every interactive component is designed in all of its states before it ships: default, hover, pressed, disabled, correct and try again.

      Motion

      Motion is feedback, never decoration. Hover a row to see its timing.

      Voice and tone

      Eight rules every screen follows

      Eight components that replaced the duplicates

      How I built it, in order

      How I use AI

      System architecture

      As the designer I needed to know where each piece of data comes from, how long it takes and what can fail. That shaped loading states, limits and copy as much as any sketch.

      What happens when you ask for writing feedback

        What the APIs taught me

        CHAPTER 09Did it work

        Learners stopped renting their German. They started owning it.

        Duolingo is excellent at bringing people back, and I never set out to beat it at that. I tested something different: what happens to confidence, motivation and commitment when learning lives in one path you paid for once. Twenty learners used Germanly for two to four weeks and were interviewed before and after.

        Before and after the pilot

        Median on a 7 point scale, n = 20, Wilcoxon signed rank

        BeforeAfter

        What predicted who kept going

        Spearman correlation with intent to continue, n = 20

        Before the pilot, learners predicted

        A subscription would keep them going better than paying once. Median 5 versus 3, p = .004.

        After living with one payment
        3of 7

        Agreement that a subscription would have kept their use the same. Most said it would have made them use Germanly less.

        What the pilot shows

        Confidence and self directed motivation rose, scattered tools consolidated, and learners' pricing beliefs flipped once they lived with a single payment. Ownership, not guilt about money, tracked with wanting to continue.

        What it cannot show yet

        It was one bundled condition with no control group, so it cannot separate pricing from the single path, or either from novelty. The next study is a 2 × 2 design with usage logs over 8 to 12 weeks.

        Learners over time

        CHAPTER 10The research paper

        The study, written up for CHI.

        CHAPTER 11What I learned

        The streak was never the enemy. What it measured was.

        What is next