@Isaac McNeil great question. Mostly using Claude as the reasoning layer for: (1) lead qualification — scoring inbound form submissions against a custom rubric and routing hot leads to humans, (2) email triage — classifying intent and drafting context-aware replies, (3) document/RAG-style Q&A on internal SOPs, and (4) multi-step decision trees where if/else logic in n8n alone gets ugly fast. n8n still owns triggers, integrations, retries, and any deterministic step. The split is basically: n8n = workflow + state, Claude = anything that needs judgment. Happy to share the exact subagent patterns and prompts I use if helpful — just DM me and I'll send them over.