Teaching our assistant to text like a person

Ask a chatbot “do you ship to Canada?” and you can usually tell what you're talking to before you finish reading: 150 words, three bullet points, a bolded header, and a closing “Let me know if you have any other questions!”. No human support agent has ever typed that. We spent a stretch this summer making Replyflow's replies read like a teammate texting, and it turned out to be less about clever prompting and more about mechanics.

Bubbles, not blocks

People send three short messages, not one paragraph. Replyflow now splits replies into separate bubbles at natural sentence boundaries — “Sure!”, then “We're open till 6 today.”, then “Want me to book you in?” — with a brief typing indicator between them. Markdown never survives the trip: no bullets, no bold, no headers in a chat bubble, ever. A comparison gets folded into a sentence, the way a person would type it.

A literal banned-phrase list

Telling a model to “sound friendly and casual” barely works. What works is banning the tells outright — “I'd be happy to help”, “Great question”, “I understand your frustration” — and enforcing it twice: once in the instructions, and again with a mechanical filter that strips the boilerplate if the model slips. Answers lead with the answer, run one or two sentences by default, ask at most one question, and mirror the visitor — if someone types “hey quick q”, they get a reply in the same register.

Your voice, not ours

The bigger unlock: paste 10–50 of your real support conversations and the assistant learns your team's actual greetings, phrasing and sign-offs — then answers in that voice instead of a stock one. Names, emails and order numbers are stripped in the process. It also uses context a person would use: it knows which page the visitor is on, and recognises returning visitors instead of asking the same questions twice.

The line we won't cross

There's a known bag of tricks for making bots pass as human — fake typing typos, stock-photo “agents” named Sarah, fake presence dots. We don't do any of it. The widget says “AI assistant” right in the header, in all 30 languages. The goal is a bot that texts like a competent teammate and hands you the conversation when a human is actually needed — not one that lies about having a pulse. We even measure it: we track how often visitors ask “am I talking to a bot?”, and work to make the answer matter less.

Try it on this page — the bubble in the corner is the assistant we're describing. Or put it on your own site in about a minute.