Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Localization rules and audit workflow for cmux UI strings, settings rows, menus, shortcuts, schema/config text, docs, command/help text, alerts, tooltips, and web messages. Use whenever changing user-facing text.
.claude/skills/manaflow-ai-cmux-localization/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | -1% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -4% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 15% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 13% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -20% | 0% |
Use this skill for any user-facing string change.
After adding or changing user-facing copy, run ./scripts/localize-changes.
The command compares the current worktree with the mainline base, discovers changed Swift localization keys and English web messages, prepares simple new or changed macOS catalog entries with minimal edits, and writes a machine-readable translation packet under git metadata. Fill value for simple translations or localization for plural or variant entries, then run the same command again. When an existing translation is still correct after an English change, fill value with that same text; the packet remembers the confirmation so later runs do not mark it for review again. Completed macOS rows are imported through scripts/localization_catalog.py merge, and scripts/localization_catalog.py check remains the final authoritative validator.
Use --base <ref> for another comparison base and --work-file <path> when a translation helper needs a visible packet. The default packet stays outside the worktree. Ambiguous catalog ownership, a key whose Swift call sites (changed or not) disagree on defaultValue, count-like new strings, unsupported Swift literal forms, and invalid generated translations stop with a concrete human-attention item so the existing placeholder, plural, bidi, omission, identity-translation, copied-English, and catalog checks stay intact.
Text(), Button(), alert titles, tooltips, menus, or dialogs.String(localized: "key.name", defaultValue: "English text"), with keys in Resources/Localizable.xcstrings. Every feature PR includes translated entries for all supported macOS languages required by scripts/localization_catalog.py (currently en, de, fr, ar, es, zh-Hant, zh-Hans, ko, ja), subject to the exact omission records below. English and Japanese entries are always required.defaultValue, English fallback text, schema descriptions, and copied English strings do not count as localization. Record deliberate invariant literals in scripts/localization-allowed-omissions.json, with the exact source and omission class. A correct translation that shares the English spelling needs a documented identityLocales exception for that key and locale; the translated entry remains required.web/i18n/routing.ts, with a matching web/messages/<locale>.json entry plus any localized data structures carrying inline translations.Before finishing a task that changes UI, Settings rows, menus, shortcut metadata, schema/config text, docs, command/help text, alerts, or tooltips:
Run ./scripts/localize-changes first and resolve every reported translation row or human-attention item.
en, de, fr, ar, es, zh-Hant, zh-Hans, ko, ja) in the feature PR, unless an exact omission record allows an absent value. Omission records still require en and ja entries.rg over changed Swift/TS/TSX/docs files for newly introduced bare English.Resources/Localizable.xcstrings, Resources/InfoPlist.xcstrings, and the linked macOS package catalogs must pass python3 scripts/localization_catalog.py check. New keys must carry all nine macOS locale entries unless covered by an exact omission record; en and ja entries remain required. Preserve printf placeholders and use plural variations for count strings where the source has a count.
Count strings are recorded in scripts/localization-plurals.json with the English source and the argument numbers that select plurals. Every required plural category must contain translated text. Use substitutions when more than one count varies or when the count is not the first argument. Arabic requires zero/one/two/few/many/other; French and Spanish include many. Keep every message inside the catalog's strings object so Xcode compiles it.
For a shared-spelling word inside a plural substitution, use an identityLocales object with reason and an explicit values list. This permits the listed leaf text (for example French %d machines) while continuing to reject an untranslated English sentence around it.
New keyboard shortcuts also need docs and Settings entries; see ../cmux-keyboard-shortcuts/SKILL.md.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 13,956 | 10,758 | -23% | 1 | 1 | 0% | 2,276 | 1,381 | -39% | 0 | 0 | — |
case-02 | fail→fail | 3,902 | 11,762 | +201% | 1 | 1 | 0% | 402 | 1,329 | +231% | 0 | 0 | — |
case-03 | fail→fail | 41,037 | 17,000 | -59% | 1 | 1 | 0% | 2,889 | 1,341 | -54% | 0 | 0 | — |
case-04 | fail→pass | 17,042 | 7,427 | -56% | 1 | 1 | 0% | 2,283 | 2,269 | -1% | 0 | 0 | — |
case-05 | pass→fail | 22,429 | 9,165 | -59% | 1 | 1 | 0% | 1,627 | 2,334 | +43% | 0 | 0 | — |
case-06 | fail→pass | 16,042 | 6,785 | -58% | 1 | 1 | 0% | 2,231 | 2,150 | -4% | 0 | 0 | — |
case-07 | fail→pass | 16,587 | 8,884 | -46% | 1 | 1 | 0% | 2,028 | 2,339 | +15% | 0 | 0 | — |
case-08 | pass→pass | 9,868 | 5,622 | -43% | 1 | 1 | 0% | 1,269 | 1,758 | +39% | 0 | 0 | — |
case-09 | fail→pass | 18,181 | 10,008 | -45% | 1 | 1 | 0% | 2,392 | 2,695 | +13% | 0 | 0 | — |
case-10 | fail→pass | 28,060 | 5,214 | -81% | 1 | 1 | 0% | 2,193 | 1,746 | -20% | 0 | 0 | — |
case-11 | fail→pass | 16,839 | 25,428 | +51% | 1 | 1 | 0% | 2,802 | 3,495 | +25% | 0 | 0 | — |
case-12 | fail→pass | 33,293 | 4,129 | -88% | 1 | 1 | 0% | 2,410 | 1,593 | -34% | 0 | 0 | — |
case-13 | fail→pass | 12,130 | 5,114 | -58% | 1 | 1 | 0% | 1,697 | 1,682 | -1% | 0 | 0 | — |
case-14 | fail→pass | 14,089 | 5,198 | -63% | 1 | 1 | 0% | 2,009 | 1,618 | -19% | 0 | 0 | — |
case-15 | pass→pass | 14,415 | 8,168 | -43% | 1 | 1 | 0% | 1,710 | 2,292 | +34% | 0 | 0 | — |
case-16 | pass→pass | 8,724 | 15,726 | +80% | 1 | 1 | 0% | 1,102 | 2,036 | +85% | 0 | 0 | — |
case-17 | pass→pass | 6,085 | 5,771 | -5% | 1 | 1 | 0% | 672 | 1,888 | +181% | 0 | 0 | — |
case-18 | fail→fail | 15,423 | 12,357 | -20% | 1 | 1 | 0% | 2,015 | 1,311 | -35% | 0 | 0 | — |
case-19 | fail→fail | 17,996 | 13,398 | -26% | 1 | 1 | 0% | 2,520 | 3,076 | +22% | 0 | 0 | — |
case-20 | fail→pass | 11,196 | 6,119 | -45% | 1 | 1 | 0% | 1,831 | 1,941 | +6% | 0 | 0 | — |
case-21 | fail→pass | 40,227 | 13,279 | -67% | 1 | 1 | 0% | 2,503 | 2,795 | +12% | 0 | 0 | — |
case-22 | fail→pass | 22,371 | 5,524 | -75% | 1 | 1 | 0% | 1,662 | 1,896 | +14% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 18 counted toward the lift figure. The other 4 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +50 percentage points is the difference between those two pass rates over the 18 comparable cases. 3 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
| Model | Method | Date | Lift |
|---|---|---|---|
| gemini-3.6-flash | verified | 9/19/2026 | +55% |
| gemini-3.6-flash | verified | 9/9/2026 | +41% |
| gemini-3.6-flash | verified | 8/4/2026 | +35% |
Other measured skills in the registry, with their headline benchmark lift.