How to make an AI SMS auto-reply sound human

SMS is an intimate channel. People notice immediately when a reply reads like a customer-service macro. We spent a few weeks tuning Gemini prompts for the SMS Bridge auto-reply and most of what worked was removing things, not adding them.

1. Ban the tells

"As an AI", "I hope this message finds you well", "Warm regards", and em-dashes are the four fastest giveaways. We ban them explicitly in the instructions the model gets on every call.

2. Match length and register

If the inbound message is six words, the reply should be around six words. If the sender writes in lowercase with no punctuation, mirror it. One line in the prompt — "match the sender's length and register" — does more than any persona description.

3. Few-shot with the owner's real texts

Ten real messages from the account owner outperform every adjective in a persona field. We ask for them during setup and inject them into every call. Vocabulary, sign-offs, and sentence rhythm come along for free.

4. Confidence gate, not blind send

We ask the model to return a confidence score alongside the reply. Anything under the user's threshold goes to a Drafts queue for a human to approve. Auto-send is opt-in per account and defaults to off.

5. Loop protection and a kill switch

Two auto-responders can text each other forever. SMS Bridge stops auto-sending to a number after three automatic replies in an hour — later replies wait as drafts. And any contact can text STOP AI to turn off automated replies to them. Both are non-negotiable if you are texting real people.