Skill: CJK Text Wrap Audit

Type: Execution

Purpose

Turn ad-hoc CJK (Korean, Japanese, Chinese) line-break firefighting into a layered, evidence-first audit. CJK wrap bugs are a class of cross-cutting cascade defects: a missing locale-scoped word-break rule, an Ant Design / MUI / Mantine cssinjs override that wins on specificity, an over-eager text-wrap algorithm, and a translation string with no phrase boundary can each independently cause text to shatter mid-syllable or strand a single glyph at line edge.

This skill prescribes a fixed sequence of diagnostic gates so the audit produces an evidence-backed report — codepoint dump, computed-style trace, and verification matrix — before any code is changed.

When to Use

• Korean / Japanese / Chinese hero or headline text breaks mid-syllable.
• A single glyph or word is orphaned at the end or start of a line.
• The bug appears only in CJK locales while Latin locales render fine.
• A user reports an awkward break in a screenshot but the code "looks

right".

• Browser zoom (WCAG 1.4.4 Resize Text) reveals breaks that are

invisible at 100%.

• A previous "fix" used <br>, hardcoded widths, or !important and

the bug came back.

When NOT to Use

• Pure Latin or RTL wrapping issues (use a generic CSS audit instead).
• Font-loading or glyph-substitution problems (FOUT / FOIT, missing

glyphs, fallback fonts).

• Print or PDF rendering pipelines (different break algorithm).
• Plain-text terminals or markdown pre blocks (wrap is a terminal

concern, not CSS).

• Layout overflow caused by fixed widths or white-space: nowrap,

not by wrapping behavior.

Inputs Required

Do not run this skill without:

• ] Screenshot of the offending render at the actual breakpoint.
• ] Locale code and the raw translation string (file path + key).
• ] Path to the component file containing the element.
• ] Live URL or deterministic repro steps.
• ] CSS / component framework in use (Tailwind, Ant Design, MUI,

Mantine, Chakra, plain CSS, …).

• ] Browser and zoom level where the bug reproduces.

Optional but recommended:

• ] List of all CJK locales the project ships.
• ] Light/dark mode information if the affected text changes color.
• ] Prior commit(s) that touched the same string or component.

Without asking the user (unless unavailable), gather these inputs directly from the repository and the running app.

Output Format

Produce a structured 7-section report. Do not produce code changes until Section 7 is approved.

1. Symptom Report — locale, breakpoint, browser, zoom, screenshot

reference, exact observed break (quote the broken line).

1. Diagnosis — codepoint dump of the raw string, computed values

of word-break / overflow-wrap / text-wrap / hyphens / line-break, and the cascade trace identifying the winning declaration with its specificity.

1. Root Cause — name the failing layer: global CSS, component

library injector, headline balancing algorithm, or translation string itself.

1. Fix Strategy — which of the four layers to touch, in what

order, and why additional layers are or are not needed.

1. Proposed Patch — minimal concrete diffs (file path + before /

after). No speculative refactors.

1. Verification Plan — explicit matrix of zoom levels ×

breakpoints × locales × theme modes to be re-checked.

1. Approval Gate — explicit statement: "Awaiting approval before

applying patch." Do not edit code until the user confirms.

Procedure

Gate 0 – Symptom Capture

Collect screenshot, URL, locale code, breakpoint width, browser, and zoom level. If any item is missing, request it before proceeding. Quote the offending broken line verbatim from the screenshot so later gates have an unambiguous target.

Halt if no screenshot or repro is available — visual bugs cannot be audited from descriptions alone.

Gate 1 – Raw String Forensics

Read the translation file containing the offending key. Dump the target string as Unicode codepoints (e.g., U+C774 U+B3D9 U+D558 U+C138 U+C694) and annotate:

• Existing NBSP (U+00A0), zero-width joiner (U+200D), zero-width

space (U+200B), or soft hyphen (U+00AD) — any of these may already be influencing the wrap.

• Logical phrase boundaries — mark which adjacent words form a

semantic unit that should not split (e.g., Korean 어절 / Japanese 文節 / Chinese 词组).

• Latin-script substrings — these still wrap by Latin rules even

inside a CJK string.

Output: a labeled codepoint table and a phrase-boundary map.

Gate 2 – Computed Style Inspection

Read the affected component file. Then trace the computed CSS for the broken element. Document the winning declaration (with selector and specificity) for each of:

• word-break
• overflow-wrap (and legacy word-wrap)
• text-wrap (wrap / nowrap / balance / pretty)
• hyphens
• line-break

Note any white-space value other than normal — it changes wrap semantics entirely.

If the component is wrapped by a UI library (Ant Design Typography, MUI Typography, Mantine Text, Chakra Text, …), record the library's injected class and where its rule comes from.

Gate 3 – Override Tracing

For each library identified in Gate 2, locate the injected stylesheet (usually via cssinjs / emotion / stitches) and capture the rule that is in effect. Calculate its specificity.

Common offenders:

• Ant Design Typography: injects .ant-typography { word-break:

break-word } (specificity 0,1,0).

• MUI Typography: injects .MuiTypography-root { … } similarly.
• Tailwind preflight: body defaults can be overridden by any

component class.

Identify a safe override location and selector. Prefer:

• The doubled-class trick: .ant-typography.ant-typography { … }

(specificity 0,2,0) — beats the library's single-class rule without !important.

• Locale scoping with :lang(ko) on the element — a pseudo-class

adds 0,1,0 specificity (unlike :where(), which erases it) and keeps wrap overrides away from ja / zh text.

• A scoped wrapper class on the page or layout level.

Avoid:

• :where(.ant-typography) — :where() collapses specificity to

0,0,0. The override will lose every cascade contest.

• !important — wins, but pollutes future overrides and is hard to

audit.

Gate 4 – Layered Fix Proposal

Propose the minimum number of layers needed. Stop at the first layer that fully resolves the symptom; only escalate if the verification matrix in Gate 5 still has failing cells.

Layer 1 — Global cascade. Add a locale-scoped base rule:

css
/* Korean only. ja/zh have no inter-word spaces — keep-all would
   remove every break opportunity there and cause overflow. */
:lang(ko) {
  word-break: keep-all;
  overflow-wrap: break-word;
}

/* Always exempt monospace contexts. */
code, pre, kbd, samp {
  word-break: normal;
  overflow-wrap: normal;
}

The UA default (word-break: normal) allows a break between any two CJK syllables — that is what shatters Korean mid-word. keep-all forbids those intra-word breaks; Korean still wraps at spaces (어절 boundaries). Japanese and Chinese are written without spaces, so keep-all would leave them no break opportunity at all — never apply it to ja / zh. For ja, keep default breaking (optionally line-break: strict for stricter kinsoku punctuation rules) and bind phrases at Layer 4 or with a phrase segmenter such as BudouX. For zh, the default break-anywhere behavior is correct. overflow-wrap: break-word is the safety net for very long Latin tokens. :lang(ko) relies on a correct lang attribute (<html lang="ko"> or on the locale subtree) — most i18n frameworks set it; verify during Gate 2.

Layer 2 — Component-library override. If a UI library injects a stronger rule, neutralize it at the same scope:

css
.ant-typography.ant-typography:lang(ko) {
  word-break: keep-all;
  overflow-wrap: break-word;
}

(Adjust the selector for whichever library is in use. Use the doubled-class trick rather than !important, and keep the :lang(ko) scope so the override never reaches ja / zh text.)

Layer 3 — Headline-specific balancing. For hero / headline text where line-length aesthetics matter, add a balanced wrap on the specific element:

html
<h1 class="text-balance">…</h1>

text-wrap: balance (Tailwind utility text-balance, or raw CSS) distributes characters evenly across lines and prevents single-word orphans on the last line. Use text-pretty only if balance is too expensive (it's a hint, not a guarantee).

Layer 4 — Translation-level phrase binding. When a specific phrase must move as a unit (e.g., a verb + its object), insert NBSP (\u00a0) between the words in the translation file, not in JSX:

json
{
  "hero.title2": "더 스마트하게 구축하고, 더\u00a0빠르게\u00a0이동하세요"
}

Use NBSP only for short, semantically tight phrases (≤4 words). Long NBSP runs prevent any wrapping at all and will overflow on narrow screens.

Gate 5 – Verification Matrix

Build an explicit matrix and mark pass/fail for every cell:

| Axis | Values | | ----------- | ----------------------------------------------- | | Zoom | 100%, 125%, 150%, 200% (WCAG 1.4.4) | | Breakpoint | sm, md, lg, xl (whatever the project ships) | | Locale | Every CJK locale + at least one Latin control | | Theme | Light, Dark (only if rendering depends on it) |

Re-test all cells after every layer added in Gate 4. The fix is only complete when every cell passes. Latin-locale regressions count as failures even if the CJK bug is resolved.

Gate 6 – Approval Gate

Present the full report (Sections 1–6 above) plus the proposed patch. State explicitly: "Awaiting approval before applying patch." Do not edit any file until the user confirms.

After approval, apply the patch exactly as proposed and re-run Gate 5 on the live build. Report the matrix again.

Guardrails

• Never use !important to win a specificity contest. Use the

doubled-class trick (.foo.foo) instead.

• Never wrap framework overrides in :where(...). It collapses

specificity to 0 and the override will silently lose.

• Never apply word-break: keep-all to code, pre, kbd, or

samp. Always carve out monospace contexts.

• Never apply word-break: keep-all to Japanese or Chinese text.

They have no inter-word spaces, so keep-all removes every break opportunity and causes overflow. Scope the rule with :lang(ko).

• Never edit a translation file without first dumping the existing

string as Unicode codepoints. Hidden characters (NBSP, ZWJ, soft hyphens) are common and easy to miss.

• Use NBSP only for short, semantically tight phrases (≤4 words).

Long NBSP runs prevent wrapping entirely and cause overflow on narrow screens.

• Do not add hardcoded max-width or width to "fix" wrapping. It

masks the cascade bug and breaks at other breakpoints.

• Do not use <br> for wrap control. It does not survive translation

length variance and breaks the moment another locale is added.

• Verify the fix in the actual component, not in DevTools-only

edits. DevTools changes don't capture cssinjs injection order.

• Always test at ≥150% browser zoom. WCAG 1.4.4 requires text to

remain usable at 200%, and CJK breaks often only appear when zoomed.

• Always test every CJK locale the project ships, not just the

reported one. The same string can wrap differently across ko, ja, and zh.

• Confirm the fix does not regress Latin locales (en, de, fr,

es, …). German is ~30% longer and is the canonical regression case for headline wrapping.

• Do not change directory structure or extract shared utilities as

part of this audit. Stay scoped to the bug.

• Honor the project's existing styling convention. If the project

uses Tailwind utilities, prefer text-balance over inline style={{ textWrap: 'balance' }}.

Failure Patterns

Common bad outputs:

• Applying word-break: keep-all globally without exempting

code / pre — breaks code blocks across the entire app.

• Applying word-break: keep-all unscoped (e.g., on body) in a

product that also ships ja / zh — those locales have no spaces, so whole clauses become unbreakable and overflow narrow screens.

• Wrapping the override in :where(.ant-typography) — specificity

drops to 0, the library rule wins, and the "fix" silently has no effect.

• Adding text-wrap: balance without fixing the underlying cascade

— balances the broken text into a different broken layout.

• Inserting NBSP between every word in a string — text refuses to

wrap on narrow screens, causing horizontal overflow.

• Editing the JSX but not the Korean translation file (or only the

English one) — the bug persists in the actual locale that has the problem.

• Testing only at 100% zoom — ships a fix that breaks at 150% or

200%.

• Using <br> to force a break — works once, breaks the moment copy

changes or another locale is added.

• Hardcoding a max-width to dodge the issue — masks the cascade

bug and reintroduces overflow at unrelated breakpoints.

• Treating a single-glyph orphan as "by design" without first

checking the cascade for a missing rule.

• Skipping the codepoint dump in Gate 1 — missing pre-existing NBSP

or ZWJ that would have explained the symptom in seconds.

Example 1 (Minimal Context)

Input:

A single Korean <h1> in a plain Tailwind project breaks as 이동하세 / 요 at the md breakpoint. No UI library is in use. Project has no global word-break rule.

Output:

1. Symptom Report — locale ko, breakpoint md (768px), Chrome

123, zoom 100%, line broken as 이동하세 / 요.

1. Diagnosis — codepoints U+C774 U+B3D9 U+D558 U+C138 U+C694

(no hidden characters). Computed word-break: normal (UA default). Cascade winner: user-agent stylesheet.

1. Root Cause — Layer 1 (global cascade). UA default

word-break: normal allows breaks between any two CJK syllables.

1. Fix Strategy — Layer 1 only. No library override or balancing

needed.

1. Proposed Patch — add to src/styles/globals.css:

css :lang(ko) { word-break: keep-all; overflow-wrap: break-word; } code, pre, kbd, samp { word-break: normal; overflow-wrap: normal; }

1. Verification Plan — re-test at 100/125/150/200% × sm/md/lg ×

ko and en.

1. Approval Gate — Awaiting approval before applying patch.

Example 2 (Realistic Scenario)

Input:

A SaaS admin portal's Korean login hero shows three rotating slogans built with Ant Design <Typography.Title>. Slogan 2 reads "더 스마트하게 구축하고, 더 빠르게 이동하세요" and breaks as "이동하세" / "요" at lg. Project uses Tailwind 3 + Ant Design 5. Hero element is src/pages/login/index.tsx. Translation key login.hero.title2 lives in public/locales/ko/common.json.

Output:

1. Symptom Report — locale ko, breakpoint lg (1024px), Chrome

123, zoom 100% and 150%, line broken as 이동하세 / 요. Bug reproduces only in ko; en renders fine.

1. Diagnosis — codepoint dump of the raw string contains no

NBSP / ZWJ / soft hyphens. Computed style on the <h1> shows word-break: break-word winning from .ant-typography (specificity 0,1,0), overriding the project's body { word-break: keep-all } (specificity 0,0,1). text-wrap is wrap (default).

1. Root Cause — Layers 1 + 2 + 3 + 4 are all required:

• Global rule was unscoped (body) and missing the monospace

exemption; it must move to :lang(ko).

• Ant Design's cssinjs rule wins over body.
• Even after the override, text-wrap: wrap will orphan 더 on

the previous line because the natural break lands one word short.

• The phrase 더 빠르게 이동하세요 is semantically a single unit

and must move together.

1. Fix Strategy — apply all four layers in order; verify after

each.

1. Proposed Patch:

src/index.css (replaces the existing unscoped body rule): css /* Korean-scoped text wrapping */ :lang(ko) { word-break: keep-all; overflow-wrap: break-word; } code, pre, kbd, samp { word-break: normal; overflow-wrap: normal; } /* Ant Design override — doubled-class beats single-class without !important */ .ant-typography.ant-typography:lang(ko) { word-break: keep-all; overflow-wrap: break-word; }

src/pages/login/index.tsx (hero <Title>, three places): tsx <Title level={1} className="!mb-3 text-4xl md:text-5xl lg:text-6xl text-balance" style={{ color: 'var(--color-heading)' }} >

public/locales/ko/common.json: diff

• "title2": "더 스마트하게 구축하고, 더 빠르게 이동하세요"

+ "title2": "더 스마트하게 구축하고, 더\u00a0빠르게\u00a0이동하세요"

1. Verification Plan — matrix:

| Cell | Result | | ----------------------------- | --------- | | 100% / lg / ko / dark | pass | | 125% / lg / ko / dark | pass | | 150% / lg / ko / dark | pass | | 200% / lg / ko / dark | pass | | 100% / md / ko / dark | pass | | 100% / xl / ko / light | pass | | 100% / lg / en / dark (control) | pass |

1. Approval Gate — Awaiting approval before applying patch.

Notes

FAST MODE (only if explicitly requested):

• Skip Gate 5's full matrix. Test only the reported zoom and

breakpoint, plus 150% as a WCAG sanity check.

• Still produce Gates 1–4. Diagnosis without verification is

acceptable in fast mode; fixes without diagnosis are not.