Agents · GuideOpen my Instagram inbox on any morning and the same questions are already waiting.
How much is the course? What's inside? What will I actually learn? Can we collaborate? Where do I even start with AI?
I get 100+ of these a day. Every single one deserves a real answer from a real person. And if I wrote every one myself, that would be my whole day.
The obvious fix is a bot. That's exactly where it goes wrong. A generic bot replies like a bot: stiff, long, full of "Dear customer" energy. The person messaged me. They wanted Ani, not a help desk. One robotic reply and the trust is gone before they finish the first line.
So I didn't build a bot. I cloned my voice into a new team member, and I kept the final say for myself.
01Why canned answers were never going to work
Think about hiring a new assistant. Option one: on day one you hand them your price list and say "go answer clients." Option two: they sit next to you for a month and listen to how you actually talk to people.
The first one knows your prices. The second one knows you. Canned replies are the price list. What I wanted was the month beside me.
That's why my agent runs on two separate things. A voice, learned from my real messages. And facts, taken only from my real business files. Put them together and a reply is both correct and recognizably mine.
02What my inbox looks like now
My agent lives in Groknbot. When a message comes in, it writes the draft. My words, my short sentences, native Armenian. The way I would write it.
My job is to read and approve. I approve, it sends. Until I do, nothing leaves my inbox.
In numbers: 100+ messages a day, and about 10 minutes of my time instead of the 2 to 3 hours it used to take.
But the part that matters more than any number: people get an answer that sounds like me, because it learned from me.
How it works, in four steps
No magic here. Just the right order, where each step covers the weak spot of the one before it.
First, the agent read all of my past Instagram messages. Not a few samples. The whole history.
From that, it learned how I reply. My tone. My words. My short answers. The way I explain prices and courses.
This is the step people skip, and it's the one that matters most. You can't describe a voice in one line like "be friendly." A voice lives in the replies you have already written a hundred times over to the same question.
A voice alone is dangerous. An agent can sound exactly like me and still quote the wrong price. That's worse than a generic bot, because people trust it.
So it also works from knowledge files built from huva.io: the real course information, prices, what's inside each course, and what students learn.
The rule is simple. The voice comes from my messages. The facts come from the files, never from the agent's imagination.
I connected the agent to ChatPlace the same way I would bring in a new hire: I invited it as a team member.
As a team member, it has access to all my Instagram conversations. Not a separate hookup per chat. The same access my own manager would have.
For me, this changes how you think about the whole thing. I didn't plug in a tool. I added someone to the team. That someone just happens to be AI.
The agent drafts every reply. Nothing is sent until I approve it. I approve, then it sends.
Some people will read that as a limitation: "What kind of automation needs you to sit there and click approve?" I see it the other way around. The approval step is the strongest part of the system.
No weird, embarrassing bot reply ever goes out under my name. Whoever writes to me is still talking to me, because the last word is mine. The agent does the heavy lifting. The trust stays with me.
And the replies go out in natural Armenian, not Armenian that reads like a translation.
03The prompt to clone your own voice
Want to build the same thing for your business? Start where I started: your voice. Give this prompt to your agent along with your past conversations, and it will pull out how you actually reply.
Role: You are a voice analyst. Your job is to learn exactly how I reply to messages, so an agent can draft replies that sound like me.
Goal: Build a reply style guide from my real past messages.
Context: I run [your business]. People message me about [prices, products, collaborations, general questions]. Below are my real past conversations. Study only my replies, not the questions.
[paste your past message threads here]
Output format:
1) My tone in 5 short rules
2) Words and phrases I use often
3) Words and phrases I never use
4) How long my replies usually are, and how I open and close them
5) How I explain prices and products
6) How I say no or redirect someone
7) 5 example replies in my voice for my most common questions
8) Topics where you are not sure how I would answer. List them. Do not guess.
Rules: Do not invent facts about my business. Quote my real wording where you can. Keep the guide short enough to paste as standing instructions for an agent.Paste the style guide into your agent's standing instructions, with your business facts file next to it. Then, before you connect it to real conversations, test it on old questions and compare its drafts with what you actually sent. If the difference jumps out at you, the voice isn't ready yet.
04Before you let an agent touch your DMs
Quick summary
- A generic bot replies like a bot. My agent replies like me, because it learned from my real messages.
- The voice comes from past messages. The facts come from files built from huva.io.
- In ChatPlace, the agent is a team member with access to every conversation.
- It drafts, I approve. The last word is always mine.
Want an agent like this for your business?
This isn't a trick to answer faster. It's a system that protects three things: your voice, your facts and your control.
I build agents like this for businesses through HUVA Agency. Message me on Instagram at @aiwithani. And yes, my agent will draft the reply. I'll be the one who approves it.
If you'd rather build agents like this yourself, join my Agentic AI: AI Agent Teams course. This exact agent is one we build there, step by step. Every other HUVA Academy course is at huva.io/courses.