Back to writings

Building a Content Agent That Writes Like You

4 min read

Most AI writing tools produce generic content that sounds like everyone and no one. After months of experimenting, I built a content agent that actually writes in my voice.

The breakthrough wasn't better prompts or fine-tuning. It was treating voice as a system, not a setting.

The Voice Problem

I tried every AI writing tool on the market. They all had the same issue: the output felt hollow. Technically correct, but missing the personality that makes writing memorable.

The problem runs deeper than tone. Voice includes:

  • Sentence structure patterns
  • Word choice preferences
  • How you build arguments
  • What examples you reach for
  • Your relationship with the reader

Most tools treat this as a prompt engineering problem. "Write in a conversational tone." "Be direct but friendly." These instructions are too abstract for consistent results.

Voice as Data, Not Instructions

Instead of describing my voice, I fed my agent examples of it. I collected 50 pieces of my best writing: blog posts, emails, proposals, documentation.

Then I built a voice analysis pipeline:

const analyzeVoice = async (samples: string[]) => {
  const patterns = {
    sentenceLength: calculateAverageLength(samples),
    vocabularyLevel: analyzeComplexity(samples),
    transitionWords: extractTransitions(samples),
    openingPatterns: analyzeOpenings(samples),
    closingPatterns: analyzeClosings(samples),
    exampleTypes: categorizeExamples(samples),
  };
  
  return generateVoiceProfile(patterns);
};

This gave me quantifiable voice characteristics instead of subjective descriptions.

Example-Based Learning

The real breakthrough came from showing the agent examples of input-output pairs. Not just "here's my voice," but "here's how I would transform this specific input."

I created a training set of 30 scenarios:

  • Raw research → blog post outline
  • Technical concept → client explanation
  • Feature list → marketing copy
  • Meeting notes → follow-up email

For each scenario, I included my actual response. This taught the agent not just what my voice sounds like, but how I think.

const trainingExample = {
  input: "New feature: real-time collaboration",
  context: "SaaS product announcement",
  myOutput: "Your team can now edit documents together without the usual chaos of version conflicts...",
  voiceNotes: "Started with benefit, used 'chaos' for emphasis, avoided technical jargon"
};

The Validation Pipeline

Quality control was critical. I built a three-stage validation system:

Stage 1: Voice Consistency Check

  • Sentence length distribution
  • Vocabulary complexity score
  • Transition word usage
  • Opening/closing pattern matching

Stage 2: Content Quality Review

  • Factual accuracy verification
  • Logical flow analysis
  • Supporting evidence validation
  • Call-to-action appropriateness

Stage 3: Human Review Trigger If confidence scores fall below thresholds, the content gets flagged for human review. This happens about 15% of the time.

const validateOutput = async (content: string) => {
  const voiceScore = await checkVoiceConsistency(content);
  const qualityScore = await assessContentQuality(content);
  
  if (voiceScore < 0.8 || qualityScore < 0.85) {
    return { status: "human_review", scores: { voiceScore, qualityScore } };
  }
  
  return { status: "approved", content };
};

What Actually Works

After six months of iteration, here's what makes the difference:

Concrete Examples Beat Abstract Instructions "Write like this specific piece" works better than "write conversationally."

Voice Patterns Are Measurable Track sentence length, word choice, structure patterns. If you can't measure it, you can't replicate it.

Context Matters More Than Tone The same voice sounds different in a technical guide versus a sales email. Train for specific contexts.

Validation Prevents Drift Without systematic quality checks, the agent's voice gradually shifts away from yours.

The Results

My content agent now handles:

  • First drafts of blog posts (85% approval rate)
  • Email sequences (90% approval rate)
  • Social media content (95% approval rate)
  • Technical documentation (70% approval rate, needs more work)

The key metric isn't just approval rate—it's whether people can tell the difference. In blind tests, readers correctly identified AI-generated content only 30% of the time.

Building Your Own

Start with voice analysis, not voice description. Collect 20-30 examples of your best writing. Look for patterns in structure, not just tone.

Build your training set around specific input-output scenarios. Generic examples produce generic results.

Implement validation early. It's easier to prevent voice drift than to correct it later.

Most importantly, treat this as a system, not a tool. Voice consistency requires architecture, not just better prompts.

The patterns I've outlined here are the foundation, but the implementation details make the difference. A content agent that truly writes like you isn't just about better AI—it's about better systems thinking applied to the problem of voice replication.

Get in touch if you want to discuss the technical architecture or see the full validation pipeline in action.