DEV Community

Cover image for Building AI Personas That Sound Human: Our Approach to Telegram Engagement
PersonymAi
PersonymAi

Posted on

Building AI Personas That Sound Human: Our Approach to Telegram Engagement

Dead comment sections kill Telegram channels. You post great content — zero reactions. New subscribers see silence and leave.

We solved this by building AI personas that engage like real people. Here's how.

The Problem With Generic AI Comments

Everyone has seen them: "Great post! Thanks for sharing! Very informative!"

These comments are worse than no comments at all. They scream "bot" and destroy trust.

Real Telegram chats look nothing like this. Real people write "lol", argue with each other, use slang, drop stickers, and type one-word reactions.

Our Persona Architecture

Every AI account in PersonymAI has a persistent identity:

Personality traits:

  • Writing style — formal, casual, slang-heavy, emoji addict, minimalist
  • Opinion pattern — bullish, bearish, contrarian, neutral
  • Aggression — scale of 0 to 100 (polite analyst to aggressive degen)
  • Language — strict Ukrainian, Russian, surzhyk, or mixed

Behavioral rules:

  • Each account has a unique typing speed
  • Some accounts comment early, others are late reactors
  • Some reply more than they initiate
  • Sticker usage varies (15% of comments include stickers)

Two accounts never produce the same output for the same input. Ever.

Opinion Drift

Static personas feel fake within a week. Real people change their minds.

We implemented Opinion Drift — accounts gradually shift their positions over time. A bullish account won't suddenly turn bearish overnight. Instead, sentiment shifts slowly based on market conditions and community reactions.

This creates realistic long-term behavior that's indistinguishable from organic users.

The 3-Pass Quality Pipeline

Raw AI output is never good enough. Our pipeline:

Pass 1: Generation
Our proprietary AI generates comments using the post content, channel niche, and persona profile. A 1400+ line prompt system handles persona injection and niche-specific terminology.

Pass 2: Self-Check
Two groups of 5 validation rules:

  • Group A: topic relevance, persona consistency, language correctness
  • Group B: length check, style verification, repetition detection

Pass 3: Short Enforce
Real chats have lots of ultra-short messages. We enforce that 35%+ of comments are 1-5 words. Making AI write "lol" instead of a paragraph is surprisingly hard.

The Cleanup Layer

Even after 3 passes, bot patterns leak through:

  • Unnecessary dashes and formal punctuation → removed
  • English slang in Russian comments → caught and fixed
  • ALL CAPS overuse → toned down
  • "I've been in crypto since 2021" → classic AI pattern, blocked
  • Analytical tone from a "degen" persona → detected and rewritten

Threading: The Secret Sauce

65-85% of our comments are threaded replies. Accounts don't just comment — they argue with each other, agree, joke, and create sub-discussions.

A flat wall of independent comments looks artificial. A thread where someone says "you're wrong" and gets three different responses looks real.

Real-Time Context

For crypto channels, our system integrates live market data — prices, volumes, 24h changes. Comments reference actual numbers and current events.

"BTC just broke 67k" hits different than "the market is showing positive movement."

Results

Channels using our system report:

  • 300% increase in organic engagement within the first month
  • 40% higher subscriber retention
  • Natural-looking discussions from day one

The bridge between an empty channel and a thriving community is shorter than you think.


We're always looking for feedback from the dev community. What would you want to see in an AI engagement system?

PersonymAI

Top comments (0)