Every Korean string needs a declared role: rebuilding 502 cues in English word order
The drill app's production cues had drifted into natural Korean word order — the exact thing the method forbids. A dedicated cue field in English word order (직독직해) was generated and order-verified for all 502 items, and an argumentation-moves layer (67 moves in 8 functions) was added for actually defending one's thinking.
The founding rule of this app's drill method is that a Korean→English production cue must be 직독직해 — Korean words arranged in English word order, chunked: "나는 · 백엔드 엔지니어야 · 약 6년 경력의". Reading the chunks left to right maps 1:1 onto the English sentence, which is the whole point: the learner's head is being retrained to assemble thoughts in English order.
The original reflex cards followed the rule. Then, over a week of adding banks (phrasal verbs, collocations, particles, prepositions…), the cue quietly became exampleKo — a natural Korean translation — because that field happened to exist. Natural translation is exactly right for the reveal (where you check understanding) and exactly wrong for the cue (where you produce). The user caught it immediately: "한글들이 전부 한국어 어순이야."
The fix: a field per role
The lesson generalizes: in a language-learning pipeline, every Korean string has a ROLE, and roles have opposite requirements —
- cue → English word order, chunked (
"만약 캐시가 미스나면 · 우리는 · 폴백한다 · DB로") - translation → natural Korean (reveal)
- explanation → natural, rich Korean (nuance, pitfalls)
Reusing one field for another role silently breaks the method. So cueKo became its own field: a workflow generated it for all 502 production items across 12 banks, with a per-bank verification pass whose single rule was order fidelity — if the chunks read like natural Korean (verb at the end when the English verb isn't), the item fails and gets rewritten. 502 of 502 merged with zero mismatches.
The argumentation layer
The same session closed the gap the user had named bluntly: vocabulary banks don't help you argue. A new bank of 67 argumentation moves in 8 functions — claim, support, concede-and-counter ("That's true to a point, but…"), hedging ("I'd lean toward…"), conditionals ("Suppose traffic doubles…"), tradeoff judgment ("X buys us Y at the cost of Z"), polite disagreement ("I'm not sold on…"), and prioritizing/concluding ("Net-net…") — wired into the workshop as its own chip row and merged into the daily-30 loop. With the connector bank as glue and these moves as joints, the "short sentences, logically chained" method now covers reasoning, not just naming.