We Extracted Our Brand Voice From 6 Months of Content — Then Made All 31 Agents Use It

The Problem

We had a brand voice problem that didn't look like a brand voice problem. It looked like a quality problem.

Our content team kept flagging agent-generated content as "off-brand" — technically correct but somehow wrong. Too formal. Too hedging. Too much passive voice. Too many bullet points where we'd use prose. Missing the direct, slightly irreverent tone our audience responded to.

The agents weren't broken. They were missing a spec. We'd never written down what our voice actually was — we just knew when something violated it.

The naive fix: write a brand voice document. The problem: a 4-page PDF doesn't translate into consistent AI behavior. Agents need quantified, structured constraints — not "be conversational." What does "conversational" mean in terms of average sentence length? Adverb density? First-person pronoun frequency? Rhetorical question rate?

We needed to measure our voice, not describe it.

The Solution

We ran the Brand Voice Extractor skill on 6 months of our best-performing content — the pieces we were proud of, the ones that got shared, the ones that got replies. The skill analyzes linguistic patterns and outputs a machine-readable voice profile.

Then we embedded that profile as a system prompt constraint in every agent that produces external-facing content.

The Process

Step 1: Compile the voice corpus.

We fed the skill 47 pieces of content we considered on-brand:

• 18 long-form articles
• 14 LinkedIn posts
• 8 email sequences
• 7 investor updates

Total corpus: ~180,000 words across 6 months.

Step 2: Extract the voice profile.

The agent analyzed the corpus and produced a structured JSON profile:

{
  "voice_profile": {
    "name": "PyratzLabs Brand Voice",
    "version": "1.0",
    "extracted_from": "180,000 words, 47 pieces, 2025-09-01 to 2026-03-01",

    "tone_dimensions": {
      "formality": 3.2,
      "confidence": 8.7,
      "warmth": 5.1,
      "urgency": 6.4,
      "humor": 4.8
    },

    "structural_patterns": {
      "avg_sentence_length_words": 14.2,
      "avg_paragraph_length_sentences": 2.1,
      "rhetorical_questions_per_1000_words": 3.4,
      "em_dash_usage": "high",
      "Oxford_comma": true,
      "passive_voice_rate": 0.04,
      "adverb_density": "low"
    },

    "vocabulary": {
      "preferred_terms": ["deploy", "ship", "run", "feed", "build", "extract"],
      "avoided_terms": ["utilize", "leverage", "synergy", "robust", "seamless"],
      "signature_phrases": [
        "We refused to",
        "The math is simple:",
        "Here's the dirty truth",
        "Zero [X]. Just [Y]."
      ]
    },

    "formatting": {
      "preferred_list_style": "em-dash bullets for short items, prose for arguments",
      "code_blocks": "always for commands and configs",
      "headers": "H2 for sections, H3 sparingly",
      "bold_usage": "key terms and emphasis, not decoration"
    },

    "openings": {
      "patterns": ["bold claim + immediate evidence", "problem statement + math", "counterintuitive fact"],
      "avoid": ["In today's world", "Are you tired of", "It's no secret that"]
    }
  }
}

Step 3: Generate the human-readable guide (for people) and the machine constraint (for agents).

The human-readable version became our internal brand guide. Key lines:

We write short sentences. Average 14 words. If you need a comma, consider whether you need two sentences instead.

We don't hedge. "Might," "could potentially," "it's possible that" — cut them. We either know something or we say we don't.

We use em-dashes — a lot. They're punchy. They interrupt like a good conversation should.

We never say "utilize." We say "use." We never say "leverage" unless we mean a physical lever. We never say "seamless."

Step 4: Embed the profile across all 31 agents.

# agent-system-prompt-addition.yaml
voice_constraints:
  profile: "pyratzlabs-voice-v1"
  enforce:
    - max_sentence_length: 22 words
    - passive_voice_rate: "<5%"
    - forbidden_terms: ["utilize", "leverage", "synergy", "robust", "seamless", "pain points"]
    - required_structure: "claim-first, evidence-second"
    - list_threshold: "use prose for <4 items, bullets for 4+"
  tone_target:
    confidence: 8
    formality: 3
    directness: 9

We also integrated the voice profile into our content brief factory:

# content-brief-template.yaml
brief_fields:
  - topic: string
  - target_keyword: string
  - search_intent: string
  - voice_profile: "pyratzlabs-voice-v1"  # always injected
  - word_count_target: number
  - required_sections: string[]

Every brief generated by the system automatically includes the voice profile reference, so every piece of content — regardless of which agent drafts it — starts from the same constraint set.

The Results

The two agents still occasionally flagged are our customer support bot (intentionally warmer tone) and our investor update agent (intentionally more formal). Those are correct divergences by design.

The unexpected benefit: new team members onboard to our voice in hours instead of weeks. Instead of absorbing tone through osmosis, they read the extracted profile and immediately understand what on-brand sounds like — quantifiably.

At PyratzLabs, 31 agents speak with one voice. Not because we wrote a style guide — because we measured what the voice already was, then made it a constraint.