Lingosee is a browser-based, voice-first roguelike for practicing conversational Japanese. You play through a sequence of real-life scenarios in Japan, replying to NPCs by speaking (or typing) in Japanese. Three strikes—freezing, falling back to English, or failing the scene's objective—send you back to Day 1. Asking for repetition or clarification is explicitly rewarded, not penalized: that's the core design idea.

Language apps teach recognition and call it conversation. The thing that actually breaks you in a real conversation isn't vocabulary — it's the four seconds after you didn't catch what someone said, and whether you freeze or ask them to repeat it. So Lingosee scores asking for repetition or clarification higher than a lucky correct guess, and the only real failure states are freezing or giving up — not being wrong.

• Six scenarios are currently built: airport immigration, buying a train ticket, reading a platform announcement, catching the right bus, a timed convenience-store checkout, and negotiating with a real-estate agent.
• Each scenario gives you a short objective (e.g. “get stamped in,” “buy the cheapest ticket,” “catch the bus to Kichijoji before it leaves”) and a small set of phrases that will advance the conversation.
• Voice output reads NPC lines aloud with the browser's speechSynthesis API; voice input listens via SpeechRecognition on Chrome/Edge and falls back to tap-to-reply and typed input on other browsers.
• After each day a “Sensei” debrief explains what worked and what didn't, and a review deck tracks vocabulary mastery.
• Progress (vocabulary mastery, run history) is stored locally in localStorage.

• React 19
• TanStack Start + TanStack Router
• Vite 8
• Tailwind CSS 4
• Radix UI / shadcn-style primitives
• TypeScript

There is no backend API and no LLM calls anywhere in the app. NPC dialogue, branching, and scoring are fully scripted and deterministic, driven by content objects in src/data/scenarios.ts and src/data/phrases.ts, and resolved by src/engine/npc.ts, src/engine/scoring.ts, and src/engine/sensei.ts. Voice and persistence live in src/engine/speech.ts and src/engine/storage.ts. An automated acceptance-test suite is in src/engine/acceptance.ts.

• src/components/ — screens and UI (Landing, Scene, Debrief, ReviewDeck, ResetScreen, WinScreen, HUD), plus components/ui/ for shared shadcn-style primitives.
• src/engine/ — all game logic: NPC turn resolution, scoring, the Sensei debrief, speech, persisted storage, and the acceptance suite.
• src/data/ — scenario and phrase content.
• src/routes/ — TanStack Router route definitions.

Prefer working locally? You need Node.js and npm — install with nvm.

git clone <this-repository-url>
cd <repository-name>
npm i
npm run dev

Other real scripts available:

npm run build    # production build
npm run lint     # ESLint + Prettier checks
npm run format   # Prettier formatting
npm run preview  # preview the production build

This project was built with Lovable.

Open your project in the Lovable editor and keep building.

• Ship faster: describe what you want to build and Lovable handles the code.
• Stay in sync: connect the project to GitHub and every change made in Lovable is committed straight to your repository.
• Full ownership: this code is yours. Push to your repository and your changes sync back into Lovable, ready for your next prompt.

• React 19
• TanStack Start + TanStack Router
• Vite 8
• Tailwind CSS 4
• Radix UI / shadcn-style components
• TypeScript

• 6 of a planned 10 scenarios are currently built.
• An eval suite for measuring NPC hidden-state leak rate, English-fallback rate, and register-grading accuracy is planned but not yet run.