What problem does it solve?
Cold email openers that claim "I looked at your site" need a verifiable, specific detail, but manually reading every prospect's customer-stories page does not scale, and letting a language model write the line freely produces hallucinated names and copied numbers.
Core Features & Use Cases
- Ten-path page discovery: Probes ten candidate paths (/customers, /case-studies, /testimonials, etc.) with free HEAD requests and a browser User-Agent, including a soft-404 guard that rejects redirects to the homepage.
- Model picker with verbatim gates: A locked prompt (gpt-4o-mini in Clay, nano-class reasoning model outside) returns client_name, evidence_quote, detail_phrase, and confidence as JSON; five code checks then verify the name and quote appear verbatim on the fetched page and reject CMS placeholders and investor/press/partner contexts.
- Deterministic line assembly: A script, never the model, builds the final clause ("your work with Notion on global spend management") with grounding, dangling-word, and verb checks, plus a ship_ready gate that holds low-confidence rows for human review.
- Use Case: Given a list of B2B prospect domains, produce a case_study_line per row that slots into "Saw {{case_study_line}}." in email 1, with roughly 50-60% ship-ready coverage on mixed B2B lists.
Quick Start
Give the skill a prospect domain like ramp.com and ask it to find a named customer from that company's case-study page and return the case_study_line for my cold email opener.
