Prompt-engineering 직독직해: forcing English word order in Korean output

The single feature I most wanted from the AI pipeline was 직독직해 — word-by-word Korean gloss in English word order. A Korean B1–B2 learner reading top-to-bottom should hear English syntax forming in their head, not Korean.

The naive prompt failed:

"All right, who's ready for some more category theory?"
  → "자, 누가 좀 더 범주론을 들을 준비가 됐나요?"   ← natural Korean: verb at the end

That's a translation, not a gloss. Useless for shadowing.

Revision 1: hard rules + good/bad examples

The Korean MUST follow English word order, NOT natural Korean order.
Each Korean chunk is a direct gloss of its English chunk, in source order.
Use Korean particles (~를, ~에, ~로, ~라고) to point at what comes next.
 
✗ BAD: {"en": "I think that he has been lying", "ko": "그가 거짓말해왔다고 나는 생각해"}
   ← sentence re-ordered to natural Korean = total failure.

Output improved but chunk size was still wrong — sometimes 1 word per chunk (no flow), sometimes 8 words (too coarse).

Revision 2: chunk size as sense group

Chunk = a "sense group" (한 호흡 단위), 2–5 English words typically.
Group by syntactic unit, NOT by word count:
  • Noun phrase: "the cat", "my only goal"
  • Verb phrase: "is going to", "has been lying"
  • Prepositional phrase: "with this video"
  • Short subordinate clause: "that he is smart"
NEVER split 1 word per chunk — destroys the flow.
NEVER make a chunk larger than a single clause — destroys English-order training.
Aim for 5–10 chunks per ~15-word sentence.

This is what SLA research actually recommends for shadowing at the B1–B2 level (the "sense group" idea goes back to Murphey 2001). Single words destroy prosody; whole clauses get translated into native order.

Revision 3: a real failure mode in the prompt

Adding the actual Korean output a reader would naturally pick — and labeling it ✗ BAD — moved the model further than any rule. It's easier to not do something when you've seen exactly what it looks like:

✗ BAD (the AI's default — DO NOT do this):
  {"en": "I think that he has been lying", "ko": "그가 거짓말해왔다고 나는 생각해"}

After all three revisions

"who's ready for"             → "누가 ~할 준비가 됐나요"
"some more"                   → "좀 더 많은"
"category theory?"            → "범주론에 대해?"
"You're all"                  → "여러분 모두"
"in the wrong room."          → "잘못된 방에 있어요."
"So this talk"                → "그래서 이 강연은"
"I hope"                      → "나는 바라건대"
"seems"                       → "~처럼 보인다"

Reading top-to-bottom now produces an English-shaped sentence in the learner's head. The whole feature works because the prompt is strict about what failure looks like, not just about what success looks like.

Pattern: when a model defaults to a fluent-but-wrong output, the strongest prompt addition is a labeled BAD example of the exact wrong behavior. Saying "do X" doesn't beat "don't do this specific thing that you keep doing."

Commit: ba90e00 — [feat] i18n + 직독직해 + Output quizzes + Decks + Playlist + project log