Thai Humanizer: Make AI-Written Thai Sound Native

Your agent writes Thai that reads like a native speaker wrote it: customer chats, posts, product pages, app UI, and translations from English.

AI writes Thai the way English is written: ของคุณ on every noun, ได้รับการ passives, ทำการ in front of verbs, นอกจากนี้ and อย่างไรก็ตาม at the start of every sentence, and "ขออภัยในความไม่สะดวกที่เกิดขึ้น" in every reply. Thai readers notice at once, and English humanizer skills can't fix it: the tells are Thai grammar, particles and politeness, not English vocabulary.

This skill makes your agent rewrite (or write from scratch) Thai that reads native:

• Knows the slot. A LINE OA reply needs ครับ or ค่ะ once per thought; a button needs none; a letter to a ministry should stay formal. The skill asks who is speaking and to whom, then writes in that register.
• Fixes translationese, not just words. Thai word order, dropped subjects, verbs instead of การ/ความ noun piles, spaces instead of English punctuation.
• Keeps the warmth. Removes template politeness but keeps real thanks, real apologies and soft requests (นะคะ, หน่อย, ได้ไหมครับ), so plain never turns cold. A shop still sounds like a shop (คุณลูกค้า, ทางร้าน), not a friend.
• Keeps every fact. Prices, dates, promo conditions, product names stay exact.
• Translates English into Thai structure by structure: classifiers, tense, relative clauses, passives, idioms, names, numbers, dates and UI strings, from published Thai translation research and the Microsoft and Mozilla Thai style guides.
• Checks translation files (i18n_check.py): missing keys, broken placeholders and HTML tags, ICU plurals (Thai needs only other), line breaks, final periods, length for UI fit, and years: it flags a year that isn't the source year or the source + 543, and a file that mixes พ.ศ. and ค.ศ. (--era be|ce enforces one). Exits non-zero, so it can run in CI. When the era isn't clear, the skill decides from the product's existing text and genre, or asks.
• Lint script included (thai_lint.py, Python standard library only): flags 30+ tells across translationese, wrappers, structure, buzzwords, politeness formulas, punctuation, particle spelling (ค่ะ vs คะ, นะคะ) and common misspellings (อนุญาต, สังเกต, ปรากฏ). Slot-aware (--slot ui|chat|post|article|formal), works on .txt, .md and i18n .json files, and reports a density score. Optional Royal Institute loanword check (--spelling ri).
• Sourced. Rules come from the Microsoft and Mozilla Thai style guides, the Royal Institute (ORST) spelling and spacing rules, Unicode CLDR, peer-reviewed Thai research on translated Thai and the ถูก passive, a study of ChatGPT's English-Thai translations, and Chulalongkorn's Sirindhorn Thai Language Institute writing course. Each rule is tagged as sourced or inferred, and the references list every source.

Tested, honestly. 16 cases: 12 AI-written Thai texts to rewrite (LINE OA replies, captions, product pages, emails, blog and help articles, an HR notice, a review reply, app strings) and 4 English texts to translate (app strings with placeholders, a marketing post, a support reply, a team email). Each was done with and without the skill, then judged blind by a separate model acting as a native Thai editor:

• Claude Opus 5.5: the judge preferred the skill's version in 12 of 16 (rewrite 9/12, translate 3/4).
• Claude Haiku 5.5: 11 of 16 (rewrite 7/12, translate 4/4).
• Every number, name and placeholder was kept in both conditions.
• Results vary between runs; an earlier run of the same set gave Opus 10 of 16. The clearest gain is in translation and app strings, where plain prompting leaves the most translationese.

Example (LINE OA reply):

• Before: "ขอบคุณสำหรับการติดต่อเข้ามาค่ะ ทางเราขออภัยในความไม่สะดวกที่เกิดขึ้น และจะดำเนินการตรวจสอบคำสั่งซื้อของคุณโดยเร็วที่สุด"
• After: "สวัสดีค่ะ ขอโทษที่ทำให้รอนะคะ 🙏 เดี๋ยวแอดเช็กออเดอร์ให้เลย ได้ผลแล้วรีบแจ้งค่ะ"
• Native is not flat: the skill keeps greetings, warmth and selling energy, and never adds a claim the source didn't make.

Example (app UI): "คุณแน่ใจหรือไม่ว่าต้องการลบรายการนี้?" → "ลบรายการนี้ใช่ไหม"

Good for: Thai shops and brands using AI for chat and posts, agencies localising English copy, developers shipping Thai app strings, and anyone whose Thai "แปลมาแข็ง".

What's inside

• SKILL.md: procedure, the top tells table, register by slot, before/after examples
• references/tells.md: full catalogue with sources and rewrites by slot
• references/register.md: particles, pronouns, brand persona checklist, formal letter frame
• references/orthography.md: spacing, punctuation, loanword spelling, common misspellings
• references/translation.md: English to Thai, structure by structure
• references/i18n.md: plurals, placeholders, Buddhist calendar, number and date formats, wrapping, fonts
• scripts/thai_lint.py: the lint (stdlib Python 3.8+)
• scripts/i18n_check.py: the translation file checker (stdlib Python 3.8+)