GOT IT! - THANK YOU (I learned so much from Nate and the community). Earlier I asked here how to clone the sales person. You said to feed it real examples, not just instructions, to log the AI draft next to what I actually sent, to review the edits weekly, and to keep a human approving. I took all of it. The stack is vCita for CRM and inbox, Twilio texting on a toll-free number that lands in that same inbox, self-hosted n8n, Claude through the API, and Google Sheets. I split the AI into two jobs. One reads customer provided data and our internal price rates. The other writes the words. Prices stay behind my approval, and the writing can earn trust faster, so they stay separate. To build the voice, I pulled about 700 messages I typed myself in vCita over 90 days, with automated ones filtered out. The AI sorted them by situation, such as first contact, price pushback and “is this legit”, and I picked what sounded most like me. Those became three files: standing rules, real examples, and a phrase bank of openers and closers. One rule is that a first text to a new customer is only a greeting, and then we wait for their reply. Before drafting, it reads the customer’s full history, so a returning customer gets a welcome back and nobody hears the same opener twice. How it runs: A new estimate request with photos gets priced from live rates, and a draft reply lands in a staff-only note on the customer’s record. Every later email or text triggers n8n. It waits two minutes so a burst of texts gets one suggestion, checks that nobody already replied, then updates one “AI suggested reply” note per customer. Nothing sends automatically. I copy, edit and send. The AI doesn’t learn on its own. Every suggestion is logged to a Google Sheet. I’m building a weekly review next that will compare each suggestion with what I sent and propose a few changes to the voice files for me to approve. I tested it by replaying 8 real recent messages. The first pass was wordy and formal next to my short replies. After tightening, it got close. It still runs a little long, and the weekly review is there to fix that.