Brand voice คลุมเครือ = drift
"Friendly, professional, helpful" ไม่ได้กำหนดอะไรเลย AI จะเฉลี่ยเป็น tone retail ไทยทั่วไป ตึงด้วย pattern ที่จับต้องได้
สิ่งที่ work
- Forbidden phrases — คำที่ไม่อยากใช้เลย "100% guarantee", "rush", สัญญาเกินจริง AI ถูกบอกให้หลีกเลี่ยง
- Required phrases — signature opener / closer / acknowledgement เช่น "ขออนุญาตเช็คให้นะคะ" ก่อน factual claim AI ใส่เมื่อ relevant
- Sign-off — close consistent หนึ่งแบบ "ทีม Klin Skin" / "Klin Skin team" สร้าง brand recall
- Custom instructions — free-form behavior ที่ไม่ครอบคลุมข้างบน ตัวอย่างที่ใช้ได้:
- "Always mention free shipping for orders ≥ ฿800." - "If a customer asks about a competitor, redirect without naming the competitor." - "When the customer is short with us, mirror their length — don't pad."
สิ่งที่ไม่ work
- adjective ลอย ๆ ไม่มีตัวอย่าง ("warm but not gushing" — gushing เป็นยังไงในมุม LLM?)
- tone description เป็นภาษาอังกฤษเมื่อ brand เป็นไทย ใช้ตัวอย่างไทย
- กฎที่ขัดกัน "Be playful" + "be formal" → เลือก random ทุกรอบ
Brand voice ตั้งต้น 5 นาที
Voice: friendly Language: Thai Formality: polite (ค่ะ/ครับ) Emoji: none Sign-off: ทีม [Brand] Required phrases: ขอบคุณค่ะ, ขออนุญาตเช็คให้นะคะ Forbidden phrases: 100% แน่นอน, rush, สบายใจได้ Custom instructions: Always offer free shipping mention for orders ≥ ฿800. When a customer apologizes or thanks effusively, mirror briefly and move on.
พอควบคุม 90% ของ reply เริ่มจากตรงนี้ ปรับต่อ