I'm a student learning AI integration and automation by building one real workflow every day for 60 days. Everything gets documented and pushed to GitHub. Starting today.
What I learned: you can use an LLM to drive actual routing logic in a workflow — not just generate text. Claude reads each email and returns a full JSON decision object that n8n acts on instantly.
How it works:
Gmail triggers n8n on every new email
Claude API classifies it — urgency, category, sentiment, summary & suggested reply
Switch node routes it:
- Urgent → instant Slack alert
- Standard → Airtable + Slack
- FAQ → auto-replied by Claude, zero human touch
Biggest lesson: getting Claude to return consistent, parseable JSON every time is harder than it sounds — and if it breaks, everything breaks. Prompt structure matters more than I expected.
Stack: n8n · Claude API · Gmail · Airtable · Slack
🔗 https://github.com/mbuguacessy-glitch/ai-support-ticket-router
59 builds to go!