Problem: Inbox Becomes a 'Warzone'
Every morning, opening the mailbox meant dealing with a jumble of order confirmations, shipping inquiries, and inventory questions from domestic trade partners and overseas buyers. Same content mixed with urgent messages, requiring manual priority sorting each time.
"This task alone took 5 hours every day. And mistakes happened often."
Turning Point: Meeting Claude Code
A team member in the Dream Team suggested, "What if we try email automation with Claude Code?"
[Before]
• Manually opening each email from inbox
• Reading content to judge category (order/inquiry/shipping/urgent)
• Manual labeling or folder moving
• Flagging important emails
• Time required: 5 hours per day, error rate roughly 8%
[After]
• Automatic inbox scan at 6 AM daily
• Claude Code analyzes sender, subject, body text
• Auto-labels by category (color-coded)
• Urgency scoring and priority marking
• Auto-generates daily summary (recorded to spreadsheet)
• Time required: 15 minutes, error rate 0.2%
Building Process: 72 Hours on Record
Day 1: Defining Rules
We clarified classification standards for each email type.
Rule set:
1. Order emails: Known buyer + contains "order" "purchase" "PO"
2. Inventory inquiry: Contains "stock" "inventory" "availability"
3. Shipping related: Contains "shipping" "delivery" "tracking" "urgent delivery"
4. Payment related: Contains "invoice" "payment" "bank transfer"
5. Other inquiries: All emails not matching above
6. Urgent flag: Sender in VIP buyer list or contains "ASAP" "urgent"
Day 2: Building Claude Code Agent
We wrote a Python script on the Windows device.
# Pseudo code (actual integration with Gmail API + Claude)
def email_classifier():
email_list = Gmail_API.get_recent_emails(last_24_hours)
for email in email_list:
analysis = Claude.analyze(
sender=email.from,
subject=email.subject,
body=email.body_first_300chars,
rules=RULES
)
category = analysis["category"]
urgency = analysis["urgency_score"] # 0-100
label = category_to_label(category)
color = urgency_to_color(urgency)
Gmail_API.add_label(email.id, label)
Gmail_API.set_color(email.id, color)
log_record(email, category, urgency)
generate_report()
Day 3: Testing and Fine-tuning
Initially there were errors. For example:
"Stock situation looks good" is positive but was misclassified as "inventory inquiry" due to the "stock" keyword
We enhanced Claude Code's contextual analysis.
Improvement rules:
• Don't judge by "stock" keyword alone
• Check verb forms ("stock up" vs "stock inquiry")
• Learn sender patterns (compare with history from same buyer)
Results: Week 3 Data
| Metric | Before | After | Improvement |
|--------|--------|-------|-------------|
| Daily processing time | 5 hours | 15 minutes | 95% reduction |
| Classification accuracy | 92% | 99.8% | +7.8%p |
| Team stress level | High | Low | Subjective |
| Missed urgent emails | 2-3/month | 0 | 100% resolved |
| Monthly time saved estimate | , | ~100 hours | , |
Unexpected Discovery: AI 'Read' the Tone
Something interesting happened. The same content written by Indian buyers appeared more urgent, while Southeast Asian buyers wrote more formally. Claude detected this pattern. Result:
"It wasn't just simple keyword analysis. It learned the sender's cultural communication style. Really amazing."
Future Plans: Automation + Human Judgment
Current system breakdown:
• Automated filtering (95%)
• Team member final review (5%, special cases)
This balance minimizes false positives while dramatically saving time.
Conclusion
Windows device + Claude Code + clear rule definition = automation. Non-developers can do it too.