From shadowing app to study platform: a cross-pack drill engine on mobile and web
Added a 1,956-entry base-verb pack plus four study lists, gave every pack recall/reverse/compose modes, and built AI mix + story generators — shared across mobile (Expo) and web (Next.js), with the data/SRS/backend living in one core package.
Mimi started as a YouTube-shadowing tool. This session pushed it toward a general English study platform without forking the codebase, by leaning hard on the monorepo's shared @shadow-ai/core package.
What got built
Data (one source of truth, two surfaces). A 1,956-entry base-verb pack — every common phrasal verb / verb+prep pattern, frequency-tiered (T1 daily core / T2 / T3) with a particle/sense overlay (up = completion, off = separation…) — plus four user-authored lists parsed deterministically from docs/*.md: 300 daily speaking patterns, 502 conversational phrasal verbs, 623 IT chunks, 296 code-explanation terms. All emitted as typed datasets in core via committed scripts/build-*.py, so the markdown stays the editable source and the TS is regenerable.
Multi-mode drills. Each pack is drillable three ways on the same card content and the same practice_card SRS: recall (Korean → English), reverse (English → Korean), and compose (write your own sentence, AI grades it via the existing /compose/check). One mode layer (drill-modes / DeckScreen on mobile, DeckDrill / DeckPage on web) covers every pack.
Cross-pack AI. Two new Spring endpoints: /compose/mix combines 2-4 chunks from different packs into one natural sentence (with a usedAll flag so it refuses to force a mash), and /compose/story weaves a handful of the day's chunks into a short passage you can shadow and cloze-test. A "Today" screen pools every pack into one interleaved daily session.
Both clients. Because data, SRS logic, the typed API, and the backend all live in core (the web's @/lib/* are thin re-export shims), porting to the Next.js web app was UI-only: generic drill components, eight pages, and a decks i18n namespace across all five locales (en/ko hand-written; ja/zh/es AI-drafted, flagged for native review).
What fought back
Three environment problems, none in the feature code, ate real time — all written up in docs/troubleshooting.md:
- Xcode 26 / SwiftUICore link error killed the iOS build until an
expo 56.0.8 → 56.0.12patch bump + clean pods. - Multiple Expo apps on one machine — the no-dev-client debug build silently loaded a different app's Metro bundle (a phantom
ExponentImagePickererror) and a different app's backend (login 404). Fixed by pinningRCT_jsLocationandEXPO_PUBLIC_API_URL. - A 12-agent adversarial review of the diff caught a real data bug: the phrase_500 parser corrupted one Korean translation when a nuance note wrapped to a second line. Rewrote it to anchor fields positionally.
Honesty notes
The content is largely valid but not certified: a 50-entry sample audit came back ~88% clean with ~2% genuine errors — fine for daily use, worth a full pass before a hard launch. The AI mix/story endpoints are built and unit-tested (mocked, zero cost) but need a backend redeploy to work in the apps; ja/zh/es UI strings are machine-drafted. All shipped on feat/mobile-app across four commits.