How We Codified Bilal's Writing Voice and Made 31 Agents Speak It

The Problem

I have a voice. Short sentences. Data first. Provocative framing. Cut the fluff.

When I write, people engage. When my agents wrote, people scrolled past. Same topics. Same platforms. Different energy. The content was technically correct but emotionally flat. It read like a smart intern summarizing my thoughts instead of me actually thinking.

The standard fix: write a brand voice guide. "Be conversational but professional. Use active voice. Avoid jargon." Useless. Every brand voice guide says the same nothing. LLMs trained on those guides produce the same nothing.

I needed something quantified. Not "be punchy" — how punchy? What's my average sentence length? What's my adjective-to-noun ratio? Which transition words do I actually use vs. the ones I think I use?

The Solution

Mr.Chief's Brand Voice Extractor. It reads actual content — not guidelines about content — and produces a machine-readable voice profile with quantified style markers.

We fed it 6 months of my LinkedIn posts, published articles, and Telegram messages to my team. It extracted patterns I didn't even know I had. Then we injected that profile into every content-producing agent's system prompt.

The Process (with code/config snippets)

Step 1: Corpus assembly. We pulled content from three sources:

voice_extraction:
  sources:
    - type: linkedin_posts
      count: 87
      period: "2024-06 to 2024-12"
    - type: published_articles
      count: 12
      period: "2024-06 to 2024-12"
    - type: telegram_messages
      count: 340
      period: "2024-09 to 2024-12"
      filter: "messages > 50 words"  # Skip one-liners
  total_words: ~62,000

Step 2: The extractor analyzes across multiple dimensions. Here's a simplified version of the output:

{
  "voice_profile": "bilal_pyratz_v3",
  "tone": {
    "primary": "provocative-analytical",
    "secondary": "direct-impatient",
    "avoids": ["corporate-safe", "hedge-heavy", "qualifier-laden"]
  },
  "sentence_structure": {
    "avg_length": 11.3,
    "short_ratio": 0.42,
    "fragment_usage": "frequent",
    "opener_patterns": ["statement-of-fact", "contrarian-claim", "number-first"]
  },
  "vocabulary": {
    "level": "technical-accessible",
    "jargon_density": 0.08,
    "power_words": ["ruthless", "broken", "actually", "zero", "real"],
    "banned_words": ["synergy", "leverage", "holistic", "robust", "scalable"],
    "contraction_rate": 0.73
  },
  "formatting": {
    "uses_lists": true,
    "uses_bold": "heavily",
    "uses_numbers": "in-every-post",
    "paragraph_length": "1-3 sentences",
    "uses_headers": "sparingly"
  },
  "rhetoric": {
    "comparison_frequency": "high",
    "uses_before_after": true,
    "ends_with": ["provocative-question", "one-liner-summary", "call-to-action"],
    "opens_with": ["contrarian-statement", "data-point", "problem-statement"]
  }
}

Step 3: This profile gets injected into agent system prompts:

# Agent content config
agents_using_voice:
  - peiy    # LinkedIn content
  - bill    # Blog writing
  - jack    # Design copy
  - warren  # Investor comms (modified — more formal)
voice_profile: "bilal_pyratz_v3"
voice_enforcement: strict  # Reject outputs that score < 0.7 on voice match

Step 4: Validation. We ran a blind test with 15 people who follow my LinkedIn:

Test: 10 posts — 5 written by me, 5 by agents with voice profile
Judges: 15 LinkedIn followers (mix of investors, founders, operators)
Method: "Which were written by Bilal vs. a ghostwriter?"

The Results

The biggest discovery: I use sentence fragments 42% of the time. "Not this." "Zero." "Every single time." No brand voice guide would have captured that. The extractor did because it counted.

Try It Yourself

1. Gather 50+ pieces of your own content (social posts, articles, emails)
2. Run the Brand Voice Extractor — the more content, the more accurate the profile
3. Review the output — you'll learn things about your own writing you didn't know
4. Inject the JSON profile into your content agents' system prompts
5. Run a blind test after 2 weeks — measure, don't assume

Your voice isn't what you think it is. It's what you actually write. Measure the difference.