Zhiyin: Learn Mandarin with Music

Introduction

Zhiyin (知音) began as a side project with two interconnected aims:

In practice, it was the latter that drove the former. In caring deeply about the problem I was trying to solve, I figured out the tools as I went.

I built Zhiyin entirely in public, sharing progress on Instagram, LinkedIn, 𝕏 and TikTok. For 60 days straight, I uploaded a video every evening documenting my learnings as a non-technical designer building my first mobile app, with the help of AI coding tools. In the process, I managed to validate demand early and build a feedback loop that shaped the product before it launched. Following a TestFlight beta with 1,300 testers, the app launched on the App Store in January 2026. It has since been downloaded over 4,000 times across 90+ countries, fuelled largely by viral TikTok content that has generated over 5M views.

At its core, Zhiyin helps users learn Mandarin Chinese with Music. Each song comes with synced lyrics, pinyin, and translations, supported by AI-generated cultural notes and grammar guides, as well as a built-in dictionary and flashcard system. It turns the music people already love into a pathway for expanding vocabulary and deepening fluency.

Design

The Problem

Many language learning tools are built for beginners. Intermediate Mandarin learners and overseas Chinese speakers are a different audience, with unique needs, and few resources built specifically for them.

Research & Discovery

I grew up in Singapore, where roughly 75% of the population is ethnic Chinese. I know this problem intimately as both a student of Mandarin Chinese myself, and a friend of many Chinese Singaporeans who struggle to develop their Mandarin proficiency.

When I embarked on this project, I combined years of informal observation with formal interviews across 8 Mandarin learners in Singapore, China, the UK and the US.

What I observed

From my years of informal observation, two behaviours stood out consistently across both groups.

What the interviews told me

Interestingly, when I asked learners how they were keeping up with Mandarin, many brought up Chinese music independently – I didn't once prompt it. Music was often treated as a supplement to methods they had tried and largely given up. These methods included Chinese language learning apps (HelloChinese, Duolingo, Pleco, HelloTalk and others) as well as textbooks. But both were almost universally abandoned after a certain period because they felt like "chores":

Yet regardless of whether these learners kept up with textbooks or Chinese language learning apps, they almost always continued listening to Chinese music. Some in a passive manner, others in a very serious and concerted way. The latter group would create an ad hoc workflow:

Needless to say, this is a process filled with friction. Countless apps and lots of copying and pasting, just to achieve one goal: expand one's Chinese vocabulary.

I asked a few of these users to show me how their notes app looked with these lyrics. In each case, I found that Chinese characters were put first, then pinyin, then English translation, for each respective line of the song.

This directly informed the design of the lyrics page for Zhiyin:

Design Goals

From these insights, I went into the project with three clear and interconnected goals.

Development

To build Zhiyin, I used Claude Code, running it directly in the terminal within Cursor, and Xcode to build the app and test it on my iPhone. Over time, two practices made a real difference to how the process went.

Phase One: MVP

I did not start by building the best possible version of Zhiyin. I started by building enough to find out if anyone wanted it. If the answer was no, there was no point going further. If yes, I could study how they used it and let that shape what came next.

Working on weeknights and weekends, guided by my design goals, and learning through building with Claude, I launched the MVP within two months. The TestFlight version launched with 10 hardcoded songs and a YouTube-powered playback system. YouTube required no account or subscription, unlike every music streaming service. I thought this would let me reach the widest possible audience with the lowest barrier to entry. Keeping the scope deliberately tight meant I could get something real into users' hands quickly and validate whether the core idea resonated.

Thanks to viral distribution methods (see Distribution below), it attracted 1,300 testers and plenty of feedback, which set the stage for a redesign.

What I learned

Phase Two: Rebuild & Launch

Based on what I had learned from my MVP testing, I rebuilt the app from the ground up.

Scaling the content

In the MVP, 10 hardcoded songs meant every grammar guide and artist context could be prepared in advance. Scaling to the entirety of Apple Music's Chinese catalogue meant those needed to generate dynamically: created on demand by large language models the moment a user selected a song, then saved to the database so subsequent users did not trigger a fresh API call. This required a different kind of thinking about data flows, caching, cost and latency, and it was precisely the kind of complexity I had deliberately deferred until I knew the product was worth building.

A personalised experience

The homepage was redesigned to be dynamic and personalised, surfacing recommendations based on a user's listening history alongside a Hot Tracks section showing what the broader user base was engaging with most. It was also an opportunity to do what the MVP had intentionally deferred: invest in the aesthetics. The MVP home screen used placeholder shapes where real artwork should have been, functional but deliberately unfinished. The rebuild brought in actual album artwork, artist images, and a custom gradient background. On the song page, removing the YouTube video made room for a more considered treatment: album art anchored to the top of the screen, with a blurred, scaled-up version of the same image filling the background. Something more alive than the flat red it replaced. A backend was built to support this, handling user data, personalisation signals and content delivery at scale.

The app launched on the App Store in January 2026.

Technical stack

• Frontend: Swift / SwiftUI for iOS

• Music: Apple MusicKit

• Database: Supabase

• Analytics: PostHog

• Payments: RevenueCat

• Development tools: Claude Code, Cursor, Xcode

App Demo

Distribution

Growth has been entirely organic, driven by three main channels:

TikTok is the primary growth channel, with over 5 million views across screen recordings of the app displaying lyrics of viral Chinese songs. It also functions as a form of A/B testing. Posting screen recordings of different app designs and tracking which ones get more views gives real-time signal about what users actually want to look at, and that feeds into my design and development process.

Instagram has also become an acquisition channel. I identify viral videos of people engaging with Chinese songs and leave comments with the Chinese characters, pinyin, and English translation of key lines. People who engage with those comments click through to my profile, and immediately see the link to Zhiyin in my bio and download it from there.

AI search has – unintentionally – emerged as another discovery path. When users ask Claude or ChatGPT for music-based Chinese language learning apps, Zhiyin frequently appears in the response – in several cases as the first recommendation. Combined with strong App Store search visibility for terms like "learn Chinese music," the app surfaces across multiple entry points without any spend.

Reflection

Zhiyin represents a full product cycle carried out independently, from design and development to distribution and continuous iteration. A year ago, that would not have been possible for me as a non-technical designer. AI enabled me to step beyond the surface of the interface and work across deeper layers of the product, including API integration, backend logic, playback behaviour, loading states and distribution strategy.

This shift gave me more ownership, but it also made design judgement more critical. When AI enables designers to shape more of the product, the central question becomes less about what can be built and more about what should be built. The fundamentals remain essential: understanding users, framing the right problems and iterating towards solutions that are thoughtful rather than merely possible. That is why this case study starts with – and spends a good deal of time discussing – the problem, rather than the tools themselves. Indeed, as Bryan Lawson wrote in 1980: