I built an n8n agent that reads a lead's website before writing the cold email — and refuses to write when it can't

Short summary

A developer built an n8n workflow that scrapes each lead's website before writing a cold email, enforcing grounding with a strict system prompt and a first-class refusal mechanism for thin or dead sites. The system writes drafts to Google Sheets for human review, costs ~$0.01 per 30 leads with GPT-4o-mini, and was validated with a 14-test acceptance suite covering trust-critical and nice-to-have scenarios. The key insight: teaching the model to refuse was more valuable than teaching it to write.

• •n8n workflow scrapes lead websites, then LLM writes grounded cold emails with evidence-based claims
• •Refusal is first-class output: thin/dead sites get [NEEDS-HUMAN] status, never faked
• •14-test acceptance suite validates grounding and refusal; costs ~$0.01 per 30 leads with GPT-4o-mini

Generated with AI, which can make mistakes.