Uzbek Humanizer

AI is fluent, but Latin Uzbek from models is still weirdly stiff: wrong apostrophes, English word order, broken suffixes, Siz mixed with -san, and calques nobody in Toshkent actually says.

uzbek-humanizer is an Agent Skill that loads whenever you ask your agent for Uzbek copy, then rewrites it like a native speaker would.

What it handles

• Calques and machine-translation patterns, with a banned-calque list
• Correct oʻ / gʻ orthography and apostrophe normalization
• Consistent register (Siz vs sen) and politeness
• Soft synonyms, collocations, false friends and Russian-isms (rusizmlar)
• Domain presets: product UI strings, quizzes/DTM, marketing, chat, bank/OTP SMS, consent and legal drafts, health soft UI
• Cyrillic requests and mixed-script traps

Before → After

• "Men haqiqatan ham bu ilovani yoqtiraman." → "Bu ilova menga juda yoqadi."
• "To'g'ri javobni select qiling." → "Toʻgʻri javobni belgilang."

Includes reference guides, worked examples, lint scripts (banned calques, stiffness, cadence, literalness) and an eval set that passes 200/200 goldens.

Honest bar: calque-safe, more natural drafts. Product quiz and marketing copy still benefit from a native pass, and heavy slang, legal or medical text should be treated as a draft.

Open source (MIT): https://github.com/xusnitdinov/uzbek-humanizer