Lingxu Liu

Independent project

Korean Pocket

A personal tool for learning Korean vocabulary.

Try the demo Read the case

Project overview

I started Korean Pocket to solve small frustrations in my own Korean study and see what I could build with AI. It became a mobile web app for collecting vocabulary, learning with flashcards and practising spelling through dictation. Using it on my phone helped me decide what to add, simplify and make more dependable.

My role
Product direction, design and hands-on evaluation
Timeline
May–June 2026 · September 2026
Current scope
Personal study tool

01Why I built it

Three gaps in my study routine

I started learning Korean with Duolingo after years of watching Korean shows. Alongside the course, I wanted a way to work on the words and skills that mattered to me.

  1. Gap 01

    Keep practising my own words.

    I came across unfamiliar words in lessons and shows. I wanted to save them together and keep practising them beyond the lesson where I found them.

    Answered in the appVocabulary →
  2. Gap 02

    Understand words through their origins.

    For Chinese speakers, Korean words formed from Chinese-character roots, read with Korean pronunciations. The shared characters can reveal meaning, even when the sounds differ.National Institute of Korean Language ↗ can connect unfamiliar vocabulary to familiar meanings. Showing the original characters makes that connection visible.

    Answered in the appFlashcards →
  3. Gap 03

    Write the words I can recognise.

    Watching shows helped me recognise words by sound, but I often could not spell them. I wanted to practise hearing a word and writing it from memory.

    Answered in the appDictation →

02What I built

One library, two ways to practise

All screens below are from the live app. The library holds 5,000+ vocabulary entries.

Gap 01 · Personal vocabulary

Vocabulary

Choose, find and organise your words.

Choose a starting point
Switch between word books or create personal notebooks. Each book keeps its study progress.
Find & inspect
Search in Korean, Chinese or English; filter by learning status, word type or lesson. Tap a word for its full details.
Make it yours
Save words to notebooks, mark familiar words as known, or add and edit entries with single-word or batch AI assistance.

Gaps 01 + 02 · Recall & context

Flashcards

Recall a meaning, then check the context.

Start from home
Choose due reviews or new words, in groups of 10, 15 or 20.
Recall & check
Listen and recall the meaning, mark Got it or Forgot, then check the answer. Details show the word type and origin.
Continue from the summary
Save difficult words to a notebook or send the group straight to dictation. Undo a swipe or resume an unfinished session.

Gap 03 · Spelling from sound

Dictation

Listen, write and check your spelling.

Choose the words
Select a lesson, words studied today or a date-based set—or bring over a completed flashcard group.
Choose how to answer
Type Korean in the app or write on paper. Set a playback interval or move through the words manually.
Check & revisit
Check the answers and save the session to dictation history for later reference.

How the parts connectFrom a word list to a study sessionExplore the flowCollapse
Choose your vocabulary

Word book or personal notebook

Remember the meaning
Flashcards

New words or due reviews

  1. Recall
  2. Check
  3. Rate
Write from sound
Dictation

A lesson or study set

  1. Listen
  2. Write
  3. Check

Finished flashcard group“Dictate this group”Same words, spelling practice

Scheduled reviews bring words back.Save words to a personal notebook at any time.

03Two decisions

How it grew with use

Neither change was planned at the start. Each one began with something I noticed while using the app, and led to a decision that changed how that part of it works.

Decision 01Vocabulary data

From easier entry to more dependable vocabulary.

  1. Problem

    Adding words by hand was slow.

    Manual entry met the original need: keeping the words I came across. Typing in every word and its details was the slow part.

  2. Two decisions
    A
    AI-assisted entry

    Fill in one word or a whole batch with AI, so collecting my own words takes less typing.

    B
    Prepared word books

    A friend learns from a fixed word book, which showed me another way to begin: pick a book and start studying straight away. To prepare the books, I had AI fill in the meanings the source lists didn’t include.

  3. New problem

    Both paths now relied on AI-generated data.

    Word types and origins kept needing correction. When I asked AI to check its own output, I only got another AI answer. With no outside reference, nothing could settle which answer was right.

    AI fillsAI checksno outside reference

  4. Final decision

    Use open Both are published by the National Institute of Korean Language, and both let you download their data.Standard Korean Language Dictionary · 표준국어대사전 ↗Basic Korean Dictionary · 한국어기초사전 ↗ as the reference.

    Searching for sources led me to downloadable Korean dictionaries. Both paths now match against them first. Anything without a clear match comes to me for review instead of being guessed.

    AAI-assisted entry · checked against the dictionaryBWord books · matched to the dictionary

Claude skill · Word-book matchingThe dictionary-matching workflow behind every word bookSee the stepsCollapse

Each time I add a word book or new words, the same matching rules have to be applied. I wrote them into a Claude skill, so every update follows the same steps. AI does the matching and formatting; anything it can’t match clearly comes back to me instead of being guessed.

  1. 01Lock the source

    Keep the Korean word, and the meaning from the source list, exactly as given. Every match is checked against them.

  2. 02Sort the entry

    Phrases, names and inflected forms (예쁜 → 예쁘다) aren’t dictionary headwords. They get a meaning, but no word type or origin.

  3. 03Look it up

    Search the Standard Korean Language Dictionary first, then the Basic Korean Dictionary. If neither has it, flag it for me.

  4. 04Match the meaning

    One spelling can be several words: 배 is stomach, boat or pear. Pick the entry that fits the source meaning.

  5. 05Copy & publish

    Copy word type, part of speech and origin, mark any AI-written translation, add audio, then sync. Existing progress stays.

Words I add myself go through the same dictionary-first matching, built into the app’s add-word flow.

Decision 02Review scheduling

From a fixed timetable to timing that follows each word.

Problem

My fixed timetable brought words back on schedule, but treated easy and hard words the same. Remembering a word within one session didn’t tell me how well I would remember it days later.

Decision

With AI’s help, I switched to FSRS (Free Spaced Repetition Scheduler) is an open-source scheduling algorithm. From your past ratings, it estimates how quickly you are likely to forget each word. It then schedules the next review for around the time recall would drop below a target level, such as 90%. Easy words come back less often; hard words come back sooner.Open Spaced Repetition ↗ instead of adding more interval rules of my own. Each rating now shapes when that word comes back. I can still practise the same words again on the same day without changing their schedule.

Before

Fixed timetable

Every word comes back on the same preset days, however easy or hard it is. Easy words keep returning long after I know them, which wastes reviews. Hard words wait too long between reviews, so I forget them before they return. Review time ends up spent on the wrong words.

Easy word
Hard word
Day 0Day 40

After

FSRS

FSRS keeps an estimate of each word’s memory: how stable it is and how difficult it is for me. Every rating updates that estimate, and the next review is set for when I am likely to start forgetting. Hard words come back sooner and easy words later, so reviews go where they are needed.

Easy word
Hard word
Day 0Day 40

Each dot is one review. The spacing is illustrative, not measured review data.

04My role

Turning AI’s capabilities into a tool I could use

I brought the learning context, tried the deployed app and made the final product choices. AI contributed ideas and technical explanations as well as implementation. The useful work happened between those roles.

What I broughtWhat AI contributedWhat I could do with it

Product judgment

Describe the study friction, compare ideas against my routine and decide what deserved to stay.

Options & technical exploration

Suggest approaches, explain feasibility and help structure requirements.

Turn a need into a usable workflow

Move from “save my words” to notebooks, word books and connected practice.

Critical evaluation

Question unreliable output, discuss and accept rules, and handle cases needing my judgment.

Matching & batch processing

Find candidates, apply agreed rules, translate and prepare updates.

Make a repeatable data process

Use references and explicit exceptions to maintain a growing vocabulary library.

Hands-on problem solving

Use each version on my phone, describe what feels wrong and check whether a change addresses it.

Coding & debugging

Build interfaces, diagnose issues and implement revisions.

Build, try and refine

Turn feedback from actual use into changes I could deploy and try again.

05What’s next & takeaways

A personal starting point, with room to grow

Korean Pocket now works for the way I study. My committed next step is to keep growing the vocabulary library as I learn, using the skill so every new word is checked the same way. If the rights question below can be solved, I’d test the core routine with a few learners, then add small things that make practice more fun, like studying with friends or quick card games.

The trade-off

Why it stays personal for now

The main constraint is content rights. The word books are based on textbook vocabulary lists owned by their publishers, so I can use them for my own study but cannot distribute them. Licensing each title would take time and an upfront cost that is hard to justify for a solo project, and building an original library would be a separate project. Releasing the app without the word books would remove the feature that makes it most useful to other learners. For now, I have chosen to keep a complete tool for my own use rather than release a reduced version more widely.

Easier to solve
  • App store release. AI features can be restricted in some regions, so the app may need different servers.
  • Accounts and sync. Other people would need accounts and cloud sync, with running costs that grow with each user.
What I learned · 01

The part I enjoy is the problem behind the problem.

Every fix in this project exposed something new. Faster entry led to AI data, AI data led to errors, and the errors led to dictionaries. I found I enjoy that chain: staying with a problem until I understand what is actually wrong, instead of stopping at the first fix that works. So far I have only done it for my own needs, where I could feel every problem myself. With real users, problems are less visible: they show up in behaviour and data, and many compete for the same limited time. Working out which one matters most, and why, is the part of product work I want to do next.

What I learned · 02

AI moved the hard part from building to checking.

AI could turn an idea into a working feature in minutes, and it sounded just as sure when it was wrong. Asking it to check its own word data only produced more confident answers. The real risk wasn’t the mistakes; it was how easy it became to stop checking. Now I ask “how would I know this is right?” before “can AI do this?” A model checking its own answer isn’t a real check. It needs a source the model didn’t produce, and choosing that source turned out to be a product decision, not a technical one.

Thanks for reading.

Here’s something to carry with you: a line from my pocket.

Demo · Coming to the app