The Problem Starts: 1,000 Messy Emails
Last week, the "buyer email auto-classifier bot" I built on Windows using Claude Code suddenly started throwing errors. The reason was simple, customers were sending emails in a mix of English, Chinese, and Korean.
**Before**: Manual classification of 500, 1,000 emails every morning (about 4 hours), with a misclassification rate above 15%
---
First Attempt: Naive Automation
The initial bot was designed with a simple rule: "subject keyword → category."
English subject: "Purchase Order" → ✅ Detected
Chinese subject: "采购单" → ❌ Not detected
Korean + English mix: "구매 문의 (Inquiry)" → 🤔 Partially detected
Root cause: Character encoding mismatch + multilingual tokenization errors
---
Unexpected Discovery: Claude Code's Multilingual Strength
Once I gave Claude Code precise instructions, it began automatically detecting languages and prioritizing them.
# Improved bot logic (pseudocode)
1. Normalize entire email text to UTF-8
2. Create language-specific keyword dictionaries (EN·ZH·KO support)
3. Rank by confidence score
4. When mixed languages detected, prioritize highest score
**After**: Classified 1,000 emails in 3 hours with 99% accuracy; human review time reduced to 30 minutes
---
3 Practical Tips Learned in Real Operations
1️⃣ Make Encoding Explicit from the Start
• Remove BOM when importing CSV files
• Explicitly instruct Claude Code to handle "UTF-8 processing"
2️⃣ Language-Specific Dictionaries > Single Rules
• Korean keywords: "구매", "발주", "가격표"
• Chinese keywords: "采购", "订单", "样品"
• Respect linguistic nuances in each language
3️⃣ Implement Confidence Scoring
• Auto-classify only 100% confident matches
• Tag 70, 90% range for manual review
• Feedback on misclassifications → bot retraining
---
The Team Is Smarter Now
Other bots in the Dream Team are now adopting this logic for their own multilingual automation tasks. We discovered that "language complexity" isn't a barrier to automation, it's an opportunity.
Next challenge: Upgrade the email classifier to automatically generate reply drafts.