slide-ko-polish

Slide KO Polish

Refine presentation slides created by translating English source text into Korean into natural Korean, and fix CJK line-breaking issues with CSS.

Resolve bundled scripts relative to this skill directory:

verify.sh
polish-loop.sh

Source: https://github.com/hackertaco/skill-forge/tree/main/slide-ko-polish

When to Use

When the Korean on the slides is awkward due to "translationese"
When line breaks that cut Korean word units in the middle are bothersome
When you want to keep English abbreviations/predicates (verifier, AC, replay) but make the sentences sound like natural Korean

Don't use for:

Materials originally written in Korean (a different kind of problem)
Technical documentation/API reference (accuracy takes priority over sentence flow)

Core Diagnosis — The 3 Major Causes of Awkwardness

Noun-predicate overuse — Repetition of the "X는 Y입니다" structure → Korean is a verb-centered language
"~의 / ~을" chains — Sino-Korean nouns strung together with particles → heavy and choppy
CJK line-breaking default — word-break: normalBreaks this Hangul at the character level

The 12 Major Signals of Translationese — Suspect Immediately When You See Them

Sources: Hanbit+ "Top 12 Common Translationese", National Institute of Korean Language "English-Korean Translation Translationese Sentences", Woorimal reporter Park Ye-jin.

Signal	English/Japanese original form	Natural language
① Pronoun overuse ("그것", "그", "그 말")	it, this, that	Omission or noun repetition
② Passive overuse ("~되어지다", "위치한다")	passive	Active
③ "~의" overuse	of, の	Spacing or omission
④ "~의 경우"	in case of	Omission
⑤ "~에 대하여"	about	Omit or paraphrase
⑥ "~하고 있는 중"	be -ing	Simple present tense
⑦ "~을 가지다"	have	"있다"
⑧ "~로부터"	from	"에게서", "에서"
⑨ "~을 통해"	through	"써서", "이용해"
⑩ "~에 있어"	in / for	Omission
⑪ "~기 전에"	before -ing	"~지 말고" / "~이 아니라"
⑫ Inanimate subject ("시스템은 ~한다")	inanimate subject	Human subject or imperative

Additional principles

Unpack nouns into predicates — End with a verb instead of a Sino-Korean noun + 합니다
Break "~의 / ~을" chains — If a sentence has 3 or more adverbial phrases, always split it
Keep English predicates (verifier, AC, replay) as they are — Koreanize them with particles and word order
Slides are spoken language — If it sounds natural when read aloud, it's OK
No abstract metaphors — English idioms (close, ship, land, kill) translated literally becomes ambiguous in Korean as to what it does. Replace it with a concrete verb.
- ❌ Phase 1을 **닫는다** (English close translated literally → what are you closing?)
- ✅ Phase 1을 **완성한다** (concrete)
- ❌ 이번 분기에 기능을 **태운다** (ship translated literally)
- ✅ 이번 분기에 기능을 **출시한다** / **내보낸다**
- ❌ 버그를 **죽인다** (kill translated literally)
- ✅ 버그를 **잡는다** / **고친다**

Single criterion for judgment

If it sounds awkward when read aloud, it's translationese. + If even one of the above 12 signals appears, awkwardness is confirmed.

CSS is only a safety net — manual control is the right answer for line breaks

Key realization: CSS only guarantees that "it won't break within a word." The browser doesn't know where to break. Korean slides should be broken by semantic units using  must be inserted directly.

CSS safety net (essential, apply by default)

body, .slide * {
  word-break: keep-all;        /* 어절 단위 — 자동 wrap 안전망 */
  overflow-wrap: break-word;
  line-break: strict;
}
/* 제목·짧은 본문은 balance — 마지막 줄에 단어 하나만 남는 것 방지 */
h1, .lede, .subtitle, .note { text-wrap: balance; }
/* 끊기면 안 되는 표현은 명시적으로 보호 */
.nb { white-space: nowrap; }

3 types of manual control tools

  — semantic-unit line break (used most often)
… — expressions that must never be broken (quotations, proper nouns, English predicates that disrupt flow, etc.)
max-width Generously (52ch about) — minimize automatic wrap triggers. Line breaks should happen only with  .

⛔ ` ` Forbidden break points — semantic units broken

Do not insert   right after a particle. In Korean, particle + predicate form one unit (grammatical unit).

Forbidden pattern	Severity	Why it doesn't work
`~을/를<br />`	FAIL	Object ↔ predicate separation (in Korean they are always a unit)
`~하는/한/할/된/될<br />명사`	FAIL	Modifier (relative clause) ↔ head noun separation. E.g.: `Phase 1을 완성하는<br />세 트랙`
`~이/가<br />`	WARN	Subject ↔ predicate separation
`~과/와<br />`	WARN	Noun list broken
`~에<br />`	WARN	Adverbial phrase ↔ predicate separation

Caution — ~은/는 is not automatically detected. Reason: in Korean, 는/은can be a subject particle or verb conjugation ending, so regex cannot distinguish them. E.g.:

❌ 말은 증거가 아니다 (mal + eun = subject particle → awkward)
✅ Phase 1을 닫는 세 트랙 (dat + neun = verb ending → natural)

Manual check: 은/는  If found, awkward if the preceding word is a noun, OK if it is a verb stem.

Example:

❌ AgentOS는 실행 증거와 verifier 판정, replay 기록을 따로따로 다룹니다.
✅ AgentOS는 셋을 따로 다룹니다 — 실행 증거, verifier 판정, replay 기록.
❌ AgentOS의 첫 단계가 자리를 잡았습니다.
✅ AgentOS의 첫 단계가 자리를 잡았습니다. (on one line)

✅ ` ` OK points

After sentence end (. , ! , ? )
After comma (, )
Use of dash (—) — X — Y Natural break in the form of
Before conjunction ( 하지만,  그래서)

Automatic detection: verify.sh Stage 1.5 catches the particle +  pattern.

Example of bad automatic wrap

Original: "하지만 시스템은 그 말을 믿기 전에, 증거부터 따로 들여다봐야 합니다."

❌ max-width: 38chIf left as is, the browser will:

하지만 시스템은 그 말을
믿기 전에, 증거부터 따로
들여다봐야 합니다.

"믿기 전에" is separated from "그 말을", breaking the meaning.

✅ Manual   If inserted:

하지만 시스템은 그 말을 믿기 전에,
증거부터 따로 들여다봐야 합니다.

Semantic unit (clause) preserved.

Mandatory Workflow — 4 steps (including verification gate)

Important: Do not separate polishing and verification. polish → verify is one action.

Step 1 — Extraction

WebFetch or read directly and organize the per-slide source text (title/body/bullets) as text

Step 2 — Proofreading

Apply the 11 translationese signal table. Per-slide Before/After.

Step 3 — CSS Patch (if HTML)

word-break: keep-all + text-wrap: balance +   manual line breaks +  protection

Step 4 — Automated Verification (MANDATORY)

verify.sh <slide-file.html|.md>

Checks the script runs automatically:

Stage 1 — translationese 11-signal regex scan (outputs line numbers)
Stage 1.5 — HTML structure (CSS, max-width, particle + , modifier + +noun, heavy quotation marks, abstract metaphor verbs)
Stage 2 — visual verification command output (manual check)
Stage 3 — spoken-language test guidance (manual check)
Stage 4 — LLM Fresh Review (Claude CLI call) — reviews English literal-translation metaphors, awkward expressions, and title-body correspondence with fresh context. Catches awkwardness that regex can't.

Exit code:

0 PASS — 0 Stage 1/1.5 signals + Stage 4 PASS
1 WARN — Stage 1.5 recommendations present or Stage 4 found review items
2 FAIL — 4 or more signals in Stage 1, Stage 1.5 FAIL pattern found, or missing word-break

If FAIL, always rework. Even with PASS/WARN, Stage 2/3 must be checked manually by a person.

Stage 4 Prompt (Reference)

Stage 4 of verify.sh evaluates with an LLM on the following 5 items:

Heavy quotation marks + noun phrase fragments
Abstract metaphor literal translation (literal translations of English close/ship/kill like 닫는다/태우다/죽이다/살아있다)
English word-order literal translation
Title-body message mismatch
Expressions clearly awkward to a Korean speaker

English abbreviations (verifier, AC, replay, etc.) are protected as intentional terminology. Noun phrase enumeration is accepted as a normal slide pattern.

Quick Reference — Frequently Appearing Patterns

Translationese (Before)	Naturalness (After)
"X는 Y입니다"	"X는 Y" (noun ending) or "X가 Y한다" (verbalization)
"~을 분리합니다"	"~을 따로따로 다룹니다"
"기준선을 만들었습니다"	"자리를 잡았습니다" / "토대를 다졌습니다"
"방식의 한계"	"이렇게 해서는 안 되는 이유"
"내부 상태입니다"	"안에서만 살아있는 상태입니다"
"타입화된 증거"	"타입이 박힌 증거" / "구조화된 증거"
"재시작 시 의도 손실"	"재시작하면 사라지는 의도"

Common Mistakes

word-break: keep-alltrusting only and leaving it to automatic wrap → words are preserved but meaning units aren't, so it's still awkward. Manual   must be inserted
text-wrap: pretty trust → prevents "only one word going to the next line" but can't catch meaning-unit breaks. On slides, balanceis safer
Translating even English predicates → verifier, replay, AC Fixed terms like this stay as is
4 or more bullets on one slide → can't catch your breath. Cut to 3 or fewer or split the slide
Fix only CSS and leave the text as is → line breaks improve, but translationese remains unchanged
Fix only the text and leave line breaking to automatic → good sentences get cut off in the wrong place and don't come alive
You need to nail all three (CSS safety net + natural sentences + manual  ) for it to work

Real-World Pattern

Recommended work order for presenters:

Don't write in English first; say the key point of each slide out loud in Korean, then transcribe it
If it still doesn't work, use this skill for post-editing

The best is when no trace of translation arises in the first place; this skill is the second best.

About this skill

When to Use

Core Diagnosis — The 3 Major Causes of Awkwardness

The 12 Major Signals of Translationese — Suspect Immediately When You See Them

Additional principles

Single criterion for judgment

CSS is only a safety net — manual control is the right answer for line breaks

CSS safety net (essential, apply by default)

3 types of manual control tools

⛔ `<br />` Forbidden break points — semantic units broken

✅ `<br />` OK points

Example of bad automatic wrap

Mandatory Workflow — 4 steps (including verification gate)

Step 1 — Extraction

Step 2 — Proofreading

Step 3 — CSS Patch (if HTML)

Step 4 — Automated Verification (MANDATORY)

Stage 4 Prompt (Reference)

Quick Reference — Frequently Appearing Patterns

Common Mistakes

Real-World Pattern

slide-ko-polish

About this skill

When to Use

Core Diagnosis — The 3 Major Causes of Awkwardness

The 12 Major Signals of Translationese — Suspect Immediately When You See Them

Additional principles

Single criterion for judgment

CSS is only a safety net — manual control is the right answer for line breaks

CSS safety net (essential, apply by default)

3 types of manual control tools

⛔ <br /> Forbidden break points — semantic units broken

✅ <br /> OK points

Example of bad automatic wrap

Mandatory Workflow — 4 steps (including verification gate)

Step 1 — Extraction

Step 2 — Proofreading

Step 3 — CSS Patch (if HTML)

Step 4 — Automated Verification (MANDATORY)

Stage 4 Prompt (Reference)

Quick Reference — Frequently Appearing Patterns

Common Mistakes

Real-World Pattern

⛔ `<br />` Forbidden break points — semantic units broken

✅ `<br />` OK points