English · 简体中文
Kachi teaches finance English in pairs: rally / selloff, revenue / earnings, coupon / yield, hedge / pledge. In a headline, one wrong word flips the story, so every word arrives beside the twin it gets confused with, and every question is decided by the difference.
6 levels · 238 concepts · 145 confusion pairs · 125 lessons · 1,697 prompts. Offline PWA, no account, progress stays on the device. English interface by default, 中文 one tap away. Live at kachi-game.vercel.app.
- Twin teach cards. A new pair shows up as two cards side by side: art, IPA, recorded audio and a plain caption for each, plus the one-line trick that separates them. Each pair is tagged with one of seven mix-up kinds, named so a twelve-year-old gets them: Opposites, Which way?, Which moment?, Whole vs. slice, Family, Look-alikes, Sentence buddies.
- Questions decided by the twin. Meaning, scenario and true/false items always offer the twin as a distractor, and fill-in word banks always contain it. A lesson interleaves its words (rally, selloff, slump, then round two) instead of asking about one word four times in a row. Answer options are words only: a picture next to "rally" would give the answer away.
- Eight exercise types: meaning, scenario, true/false, fill-in, match, listen, read-the-news headlines (levels 1–4) and spot-the-flaw audits (level 5).
- Research and quant practice. Level 5 has 36 original English report cases: 24 in lessons and 12 fresh cases leading the six reviews. Diagnose flawed claims, check calculations, choose validation methods and judge what the evidence supports.
- No hearts, no failing. A miss comes back once at the end of the lesson as a rehearsal (no XP, no mastery credit). Combos reward runs; the hedgehog mascot, a soft chime and a short vibration give feedback. All motion, sound and haptics follow one Effects slider.
- Spaced review. Mastery is tracked per pair on a 1, 3, 7, 14, 30, 60-day ladder. Due pairs appear on the home screen and are badged in the album.
- Three tabs. Learn is the path, with the next pair up front. Pairs is an album of every confusion pair; unmet ones show as silhouettes. Me holds your streak, the pairs you slip on most, language, effects and your data.
- Pair cards. Any pair you've met can be exported as a 1080×1440 PNG card.
Levels follow what you are reading, not which discipline a word belongs to. Every level is open; level 1 is where a new learner starts.
| Level | Pack | When it's done, you can… | Concepts |
|---|---|---|---|
| 1 | t1-finance-news |
Understand any finance app's headlines | 42 |
| 2 | t2-company-earnings |
Tell whether a quarterly release is good news | 49 |
| 3 | t3-bonds-rates |
Follow yield, auction and maturity stories | 38 |
| 4 | t4-forex-macro |
Read currency, central-bank and policy news | 37 |
| 5 | t5-quant-research |
Spot which backtest claim does not hold | 32 |
| 6 | hedge-or-pledge |
Read private-bank and structured-product material | 40 |
Every word must appear in the level's real-news corpus before it is taught
(content/corpus, enforced at build time). Course text is original: each pack
ships zh-CN and en-US course.json files, generated from source.mjs for
levels 1–5 and hand-written for level 6. The validator keeps the two locales
structurally identical. Original headline drills live in
content/packs/*/headlines.mjs; level 5's bilingual report cases live in
content/packs/t5-quant-research/audits.mjs. Their figures are hypothetical,
not quotations or actual fund results. Fill sentences for words whose corpus example
never uses the term itself live in content/packs/fill-overrides.mjs.
pnpm install
pnpm run dev # vite dev server
pnpm run test # vitest
pnpm run check # biome + content validation + tsc + vitest + vite build + build boundaries
pnpm test:e2e # production build + Playwright (desktop and mobile)
pnpm run generate:course # regenerate content/packs/*/{zh-CN,en-US}/course.json from source.mjs
pnpm run generate:art # draw missing term art (then strips the white ground)
pnpm run strip:art # make term art transparent; idempotent
pnpm run generate:icons # rebuild PNG icons from public/icon.svg (needs rsvg-convert)- Offline first. Level 1's English edition is in the main bundle (about 509 kB initial JS against a 650 kB budget), so the first screen always renders. Every other edition, including level 1 in Chinese, is a lazy chunk that the service worker precaches. If a stored Chinese preference cannot load, the app falls back to the bundled English.
- Switching is atomic. Level and language switch together only after the course has loaded. A failure keeps the current course and offers a retry inside the level picker. The chosen level and language are remembered.
- Your data. Progress lives in IndexedDB on this device. Export and import are under Me, and nothing is sent anywhere.
- Art. Term art is transparent WebP. Generated images get their white
ground removed by
scripts/strip-art-background.mjs; interior whites survive and the drop shadow becomes translucent ink. - Type. Fonts are self-hosted: Bricolage Grotesque for display, Figtree for text. Chinese uses the system CJK face.
- No dark mode, on purpose. The warm-paper, ink-outline look is the brand. Revisit only for a real night-reading use case.
- CI (
.github/workflows/ci.yml) runspnpm run checkon every push and PR. Lefthook runs biome on staged files.
Every document leads in English, with a Chinese edition beside it.
| English | 中文 | What it covers |
|---|---|---|
| docs/vocabulary-system.md | docs/vocabulary-system.zh-CN.md | The six tiers, confusion-pair types, cross-pack deduplication, the two authoring gates |
| docs/backlog.md | docs/backlog.zh-CN.md | What is not built yet, and what is deliberately not being built |
- Code: AGPL-3.0-only.
- Original course and teaching content (
content/): CC-BY-SA-4.0. - Third-party news quotations, links and brands keep their own rights and are not relicensed with the course pack.