i18n sweep
UI grows faster than translations. Typed catalogs catch missing keys, but nothing catches hardcoded strings that never became keys — those slip past every typecheck and show up as mixed-language screenshots. This skill hunts both.
Step 1 — Map the i18n system
Detect how this project translates: typed message catalogs, i18next/react-intl, or a custom dictionary. Find:
- The locale list (e.g.
en, uz, ru, ko) and the catalog file per locale.
- The lookup mechanism (
t('key'), typed dict access, component wrappers).
- The reference locale (usually the most complete one — verify, don't assume it's
en).
- Any documented exceptions in
AGENTS.md, CLAUDE.md, client-specific project instructions, or decision docs—some content is intentionally untranslated (for example, practice text, code snippets, or brand names). Respect those.
Step 2 — Find the gaps (two different hunts)
Hunt A — catalog drift. Diff every locale's keys against the union of all keys. Report missing keys per locale. If catalogs are typed, tsc output helps but is not sufficient — check for keys whose value is an empty string, a TODO, or a copy of another language.
Hunt B — hardcoded strings. Scan the scoped source (JSX text nodes, Text components, placeholder/title/aria-label/alt attributes, toast/error messages, empty states, dialog buttons) for user-facing string literals that bypass the i18n system. Heuristics: literal strings containing spaces or unicode letters in render code; template literals building sentences. Exclude: log lines, test files, config, CSS values. List every hit with file:line.
Report counts before fixing: "N missing keys across M locales, K hardcoded strings" — and if the sweep was scoped, say what was NOT scanned.
Step 3 — Fix
- Convert each hardcoded string into a key + catalog entries for every locale, following the project's existing key-naming convention.
- Fill missing translations with native-quality text, matching the tone of that locale's existing entries (formal vs casual register matters — mirror what's there).
- Keep interpolation variables/placeholders identical across locales; never translate variable names.
- Preserve pluralization structures if the framework supports them.
- Don't invent locale-specific formatting: reuse the project's existing date/number formatting helpers.
Step 4 — Verify
- Typecheck + build.
- Re-run Hunt A: zero key drift.
- Re-run Hunt B on the touched files: zero remaining user-facing literals (minus documented exceptions).
- If a dev server/preview is available: render the worst-affected pages in each locale and confirm no fallback language bleeds through, and no layout breakage from longer strings (German/Russian/Uzbek text is often 30-40% longer than English).
Report
Per-locale coverage table (keys present / total), list of files touched, strings converted, translations added, documented exceptions skipped, and anything that needs a human native-speaker review. Don't commit/push unless asked (see the ship skill).
1---2name: i18n-sweep3description: Sweeps a codebase for internationalization gaps, including hardcoded user-facing strings, missing translation keys, and locales that drifted out of sync, then fills the gaps with proper translations and verifies the result. Use when UI is untranslated, screenshots show mixed languages, the user asks to add a language, or an i18n audit is requested for a page, directory, or whole project.4---56# i18n sweep78UI grows faster than translations. Typed catalogs catch *missing keys*, but nothing catches *hardcoded strings that never became keys* — those slip past every typecheck and show up as mixed-language screenshots. This skill hunts both.910## Step 1 — Map the i18n system1112Detect how this project translates: typed message catalogs, i18next/react-intl, or a custom dictionary. Find:13- The locale list (e.g. `en`, `uz`, `ru`, `ko`) and the catalog file per locale.14- The lookup mechanism (`t('key')`, typed dict access, component wrappers).15- The reference locale (usually the most complete one — verify, don't assume it's `en`).16- Any documented exceptions in `AGENTS.md`, `CLAUDE.md`, client-specific project instructions, or decision docs—some content is intentionally untranslated (for example, practice text, code snippets, or brand names). Respect those.1718## Step 2 — Find the gaps (two different hunts)1920**Hunt A — catalog drift.** Diff every locale's keys against the union of all keys. Report missing keys per locale. If catalogs are typed, `tsc` output helps but is not sufficient — check for keys whose value is an empty string, a TODO, or a copy of another language.2122**Hunt B — hardcoded strings.** Scan the scoped source (JSX text nodes, `Text` components, `placeholder`/`title`/`aria-label`/`alt` attributes, toast/error messages, empty states, dialog buttons) for user-facing string literals that bypass the i18n system. Heuristics: literal strings containing spaces or unicode letters in render code; template literals building sentences. Exclude: log lines, test files, config, CSS values. List every hit with `file:line`.2324Report counts before fixing: "N missing keys across M locales, K hardcoded strings" — and if the sweep was scoped, say what was NOT scanned.2526## Step 3 — Fix2728- Convert each hardcoded string into a key + catalog entries for **every** locale, following the project's existing key-naming convention.29- Fill missing translations with native-quality text, matching the tone of that locale's existing entries (formal vs casual register matters — mirror what's there).30- Keep interpolation variables/placeholders identical across locales; never translate variable names.31- Preserve pluralization structures if the framework supports them.32- Don't invent locale-specific formatting: reuse the project's existing date/number formatting helpers.3334## Step 4 — Verify35361. Typecheck + build.372. Re-run Hunt A: zero key drift.383. Re-run Hunt B on the touched files: zero remaining user-facing literals (minus documented exceptions).394. If a dev server/preview is available: render the worst-affected pages in each locale and confirm no fallback language bleeds through, and no layout breakage from longer strings (German/Russian/Uzbek text is often 30-40% longer than English).4041## Report4243Per-locale coverage table (keys present / total), list of files touched, strings converted, translations added, documented exceptions skipped, and anything that needs a human native-speaker review. Don't commit/push unless asked (see the `ship` skill).
abdurshd