Build your own AI Korean tutor with Claude Code and Inworld

npm install -g @inworld/cli
inworld login
inworld workspace select
claude mcp add inworld -- inworld mcp

mkdir -p ~/projects/kori
cd ~/projects/kori
claude

Build a working Korean–English language-learning web app called Kori.

Work autonomously using sensible defaults. Inspect the existing project first and reuse its stack and patterns. If the folder is empty, create a lightweight responsive app with a Node.js backend. Implement and test the app through its normal user flows.

PRODUCT
Kori helps English-speaking beginners learn Korean through short lessons and live spoken conversations with Nari, an encouraging AI tutor.
Brand header: "kori/ Language App"
Main heading: "Your Korean AI Tutor"
Supporting text: "A little listening, a little speaking. Make Korean part of your day."

LESSONS
Create three selectable chapters with three phrases each.
1. First meetings
- 만나서 반가워요. — Nice to meet you.
- 안녕하세요. — Hello.
- 저는 학생이에요. — I'm a student.
2. At the café
- 커피 한 잔 주세요. — One coffee, please.
- 얼마예요? — How much is it?
- 감사합니다. — Thank you.
3. Getting around
- 역이 어디예요? — Where is the station?
- 여기서 내려요. — I get off here.
- 다시 말해 주세요. — Please say that again.
Each phrase should show Korean, accurate romanization, and English meaning. Include previous/next controls, phrase playback, a "Mark learned" toggle, and saved progress using localStorage.

LIVE AI TUTOR
The main learning interaction must be a real-time spoken conversation.
- Start button: "Start live conversation"
- Stop button: "End session"
- No typed reply field, Send button, or manual record-and-submit controls.
- Nari speaks first and introduces the selected phrase.
- The learner answers aloud.
- Detect turns automatically and support interrupting Nari.
- Nari responds aloud, gives gentle corrections, and continues a simple role-play.
- Allow the learner to speak Korean or ask questions in English.
- Keep each tutor turn short, usually one or two sentences.
- Explain one correction at a time and give the learner another chance.
- Accept natural variations and explain casual versus polite language accurately.
- Never claim to assess pronunciation from a text transcript alone.
Example:
If the target is "만나서 반가워요" and the learner says "만나서 반가워," explain briefly that adding 요 makes it polite, model the phrase, and invite the learner to repeat it.

INWORLD INTEGRATION
Use Inworld for the live voice experience.
Intended configuration:
- Tutor model: Gemma 4 26B A4B
- Previously documented model ID: inworld/models/gemma-4-26b-a4b-it
- Voice: Minji, the friendly Korean female system voice
- Speech model: inworld-tts-2
- Browser transport: Inworld Realtime API over WebRTC
Verify current model IDs, voice availability, API fields, authentication, and supported features before implementing. If something has changed, explain it and use the closest supported approach that preserves the requested experience.
Start with:
- https://docs.inworld.ai/llms.txt
- https://docs.inworld.ai/realtime/quickstart-webrtc
- https://docs.inworld.ai/portal/session-tokens
- https://docs.inworld.ai/api-reference/voiceAPI/voiceservice/list-voices
- https://github.com/inworld-ai/inworld-api-examples
Prefer the maintained official example. Read relevant documentation pages as Markdown by appending .md.
Use the same Minji voice for phrase playback and live tutoring. Keep model and voice configuration centralized.

SECURITY AND ACCESS
- Keep INWORLD_API_KEY exclusively on the backend.
- Reuse an existing server-side credential if available.
- Never print credentials or put them in frontend code, URLs, logs, or committed files.
- If credentials are missing, ask me to set them locally through a hidden prompt or a gitignored server-only environment file.
- Mint the appropriate session token on the backend for direct Realtime connections.
- Authenticate and rate-limit token routes.
- Keep development access local by default.
- Support a configured HTTPS deployment origin for mobile use.
- Do not expose the app publicly without explicit authorization.
- Avoid confusing expired-link errors. Show a useful session-expired screen and a clear recovery path.
- Provide a convenient restart command and explain any required sign-in after restart.

VISUAL DESIGN
Use this reference for the overall visual direction:
https://dribbble.com/shots/26154310-AI-Powered-Translate-and-Languages-App-Design
Create a polished, responsive interface with:
- Violet and lavender accents.
- Soft off-white backgrounds.
- Rounded cards, subtle borders, generous spacing, and readable typography.
- A chapter sidebar on desktop and a compact chapter selector on mobile.
- A prominent purple lesson card.
- Korean and English text that remains readable on small screens.

LIVE CONVERSATION SCREEN
When a conversation starts, open a dedicated voice screen.
- On desktop, use a centered rounded dialog with a softly blurred backdrop.
- On mobile, use a full-height screen.
- Use a soft lavender, pink, and pale-blue background.
- Put a luminous, iridescent orb in the center with a subtle dotted ring.
- Animate the orb according to actual connection and speaking/listening states.
- Respect reduced-motion preferences.
- Show clear states such as:
  "Connecting to Nari…"
  "Nari is speaking"
  "I'm listening…"
  "Your turn. I'm listening."
- Show a compact Korean ↔ English card with the current lesson phrase and its translation.
- Place "End session" at the bottom.
- Do not show a scrolling transcript box in the live screen.
- Avoid showing technical implementation details in the learner-facing interface.

RELIABILITY
Handle:
- Microphone permission denied.
- Insecure browser context.
- Missing or unavailable microphone.
- Blocked audio playback.
- Expired app sessions.
- Inworld authentication, quota, and network errors.
- Connection timeout or disconnection.
- Ending a session during connection setup.
- Starting another session after ending one.
Ending a session must stop microphone tracks, close WebRTC/data-channel connections, cancel pending requests, stop playback, and release resources. Clearly show when the microphone is active. Opening the voice screen must not depend on a browser pop-up window.

VERIFICATION AND DELIVERY
Verify:
- Chapter and phrase navigation.
- Progress persistence.
- Mobile layout and live-screen layout.
- Korean phrase playback with Minji.
- A real Gemma-powered tutor greeting and correction when credentials are available.
- WebRTC connection and session cleanup.
- Repeated start/end cycles.
- Useful recovery from denied permissions and expired sessions.
- No credential exposure in browser-delivered files.
Distinguish simulated-microphone tests from real-device microphone testing. Never claim that an untested feature works.
Deliver the working app, a usable local launch link, simple restart instructions, required configuration, and a concise report of what was tested and any remaining limitations.