From flashcards to a live AI interviewer — four training modes in one pass
The drill app got its missing layer: an AI mock interviewer that digs into your spoken answers, a connector-cloze chaining drill, a speed round, weak-card repair, and a 60-second speech mode — plus a fix for category drills that silently hid most of their cards behind SRS pacing.
A blunt question set the direction: "냉정하게, 카드만 달달 외우면 면접 영어가 되냐?" The honest answer was no — flashcards build the reflex layer (chunks surfacing without translating in your head), but they don't train live assembly: chaining those chunks into answers while an interviewer pushes back. So this batch added the layers above the cards.
The four modes
AI mock interview. The model plays the interviewer. One new endpoint (POST /api/practice/interview/mock) takes the dialog history and returns strict JSON {question} — an opener on an empty history (a session seed varies the topic), otherwise a follow-up that digs into the candidate's last answer: a why, a tradeoff, a what-breaks-at-scale. The phone screen drives the loop: question → answer by voice (the existing mic → Whisper-grade transcription) → the existing lenient grader → next follow-up. Five questions per session. All the pieces already existed — the transcription path, the lenient interview/check, the gemini-first complete() plumbing — so the whole feature is one prompt class, two DTOs, one service method, one screen.
Chaining. The methodology here is "drill short sentences, connect them logically" — and the connective tissue deserved its own drill. Each connector card carries a real 2-sentence example ("We deploy to staging. Once that's done, we run smoke tests."); the chaining drill blanks the connector out (___) and you speak the full chain. One regex, no new content, and it reuses the connector's SRS key so mastery tracking stays unified.
Speed round & weak-card repair. An 8-second countdown that auto-reveals the answer when it expires (pressure without trapping you), and a scope that collects only cards you've missed twice or more — the SRS data was already there (lapseCount), nobody was reading it.
60-second speech. Twenty-two explain-this topics (idempotency, N+1, circuit breakers…), a countdown, and the lenient grader pointed at a structure-focused question string — "judge only whether the core is understandable and logically connected." Zero backend change: the instruction rides inside the existing question parameter.
The bug worth writing down
The user opened the 44-item phrasal-verb deck and saw a dozen cards: "왜 카드 전체가 아니고 일부만 보이냐." Every drill was built with the SRS session builder — due cards plus at most 12 new per day. That's exactly right for the daily loop and exactly wrong for a category tile, where the intent is "let me drill this bank." Category drills now show the whole bank shuffled; only the hard-capped "오늘의 30개" keeps SRS pacing.
The general lesson: a pacing policy is a property of the entry point, not the engine. And when you cap a list, cap it loudly — a silent cap reads as "this is everything," which is worse than no cap.