How to get AI to yap like you

In 2026, everything reads like AI wrote it. Humans have started mimicking AI phrasing, and at the same time, people assume anything polished (or anything sloppy) must be generated. The line between "written by a person" and "written by a model" has disappeared — and the people receiving your messages can feel it.

The worst outcome isn't that your message sounds robotic. It's that it sounds like someone else. A cold outreach that opens with a compliment you'd never give. An objection response that's too agreeable when you'd normally push back. The recipient doesn't think "this is AI" — they think "this isn't you," and the trust is gone before you've made your point.

We see this play out in Fuchka every day. Of the 60-70% of people Fuchka connects you with, about 40% respond. Most of those are either already sold on your offer or confused about where the conversation is heading. A smaller slice suspects the message is AI-written. That suspicion alone is enough to kill a deal — not because AI is bad, but because generic is bad.

The problem

AI-generated messages fail for three reasons, and none of them are about the model being stupid:

• Tell-tale phrasing. Em dashes, "I'd love to," "hope this finds you well." The model has a voice, and everyone who uses it sounds the same. Your prospect has seen this exact sentence structure in three other DMs today.
• No situational awareness. AI can't read the room. It doesn't know when to be blunt, when to back off, or when the other person is testing you. It treats every conversation like the first one.
• Prompts that decay. Users don't know how to improve their prompt, so they tack on another instruction at the end or scrap the whole thing and start over. The prompt gets worse with every "fix."

I knew the theoretical fix for this over a year ago but felt the model and context architecture was too complex to build at the time. I didn't even know the term "meta-prompting" back then. I was just working from my own experience of bending AI to do exactly what I needed, paired with an Akinator-style questionnaire that gets users to fill in gaps they've missed or resolve contradictions the AI can spot and course-correct.

What people try (and why it doesn't work)

Setting a persona. "You are a friendly, casual sales rep who uses short sentences." This gives you a character, not your voice. It's a costume the model puts on — it'll sound like someone, just not you.

Fine-tuning. Train on your actual messages. In theory this is the gold standard. In practice you need far more data than most people have, and the moment your pitch changes or you target a new audience, the model can't adapt. You've frozen yourself in amber.

Patchfixing the prompt. The most common move: something sounds off, so you add "don't use em dashes" or "keep it under 3 sentences" at the bottom. Do this enough times and the instructions start contradicting each other. The model doesn't know which rule wins. Context rots from the inside out.

Grabbing a prompt from the internet. Someone on Twitter shares "the ultimate outreach prompt." You paste it in. It was written for their product, their audience, their voice. One-size-fits-all doesn't fit.

Stacking prompting techniques. CAPITALIZING instructions, emotional prompting ("this is very important to my career"), chain-of-thought reasoning. Each technique has merit in isolation. Layer them without knowing what you're doing and you get the same mess as patchfixing — conflicting signals, unpredictable output, and a prompt only its author can debug.

Meta-prompting alone. This one gets close. You ask a model to rewrite your prompt for you, and it does a decent job structuring things. But it writes the changes without asking why you want them. It optimizes the surface without understanding the intent. The result is cleaner but still not yours.

The fix: the Messaging Agent

The Messaging Agent combines two ideas:

1. Let the AI ask questions before it touches the prompt. When you request a change — "make it sound less formal" or "handle price objections differently" — the model doesn't just rewrite. It asks what you mean. How informal? Do you swear? Do you use slang or just drop the corporate speak? What does a good objection response look like to you? Think of it as an Akinator for your voice: pointed questions that surface what you actually want, not what the model assumes you want.

2. Rewrite the entire prompt, not just the line that changed. Every edit is a full rewrite that incorporates the new intent into the whole system. No appending, no patching, no contradictions. The prompt stays coherent because it's rebuilt from scratch each time, with every instruction aware of every other instruction.

Pro-tip: use the same model for the questionnaire and the rewrite. When one model asks the questions and the same model rewrites the prompt, the output is more cohesive — it already understands the nuance of the answers because it asked the questions in the first place.

Tutorial

Try it

The Messaging Agent is live in Fuchka. If your outreach sounds like everyone else's — or worse, like everyone else's AI — it's because the prompt was built by guessing instead of asking. The Messaging Agent is available on all paid plans. Sign up and try it — your first messages will show you the difference between AI that writes for you and AI that writes as you.