AI Systems Instructor • Real Estate Technologist
Quick Answer: Collect 5-10 writing samples that represent your best voice. Use Claude to analyze them for patterns (sentence length, tone, vocabulary, structure). Build a Context Card with voice rules, constraints, and 2-3 few-shot examples. Test across three content types. Iterate until AI output passes the 'would I actually post this?' test.
Every agent who tries AI for the first time gets the same result: polished, professional, and completely generic output that sounds like it was written by a corporate communications department. That's because AI doesn't know your voice until you teach it. The difference between agents who abandon AI and agents who build it into every workflow comes down to one thing: a structured brand voice document that makes AI output sound like them. This guide walks you through the Style Match technique—collecting writing samples, identifying your voice patterns, building a Context Card that captures those patterns, and testing it across different content types until AI consistently produces output in your voice.
Tools Needed
Claude or ChatGPT, 5-10 samples of your past writing (emails, social posts, listing descriptions), a text editor or notes app
Gather 5-10 pieces of content you've written that sound most like you at your best. These could be client emails, listing descriptions, social media posts, newsletter sections, or even text messages to clients. The key is authenticity—choose content where you were genuinely communicating, not content where you were trying to sound like someone else. Avoid anything written by an assistant, a marketing company, or copied from a template. Include a mix of content types: at least one email, one social post, and one longer piece (listing description or blog section). The more variety in format, the more accurately AI can extract the patterns that are consistent across everything you write. Save these in a single document labeled 'Voice Samples.' This is the raw material your brand voice will be built from.
Tip: If you struggle to find writing you're proud of, look at your sent email folder. The emails you write when you're being honest and direct—not performing—are usually your most authentic voice. Client check-ins, market updates to your sphere, and post-closing follow-ups are gold mines.
Paste your writing samples into Claude and use this prompt: 'Analyze these writing samples from a real estate professional. Identify patterns in: (1) average sentence length, (2) tone and formality level, (3) vocabulary preferences and frequently used words, (4) structural patterns (how they open, transition, and close), (5) what they never do (formality they avoid, phrases they don't use), (6) personality traits that come through in the writing.' Claude will return a detailed voice analysis. This is the Style Match technique in action—instead of describing your voice from memory (which is unreliable), you're extracting it from evidence. Review Claude's analysis and correct anything that doesn't feel right. If it says you're 'formal and measured' but you think of yourself as casual, look at your samples again. The samples don't lie, but the analysis might miss nuance. According to Anthropic's prompt engineering guide, providing concrete examples consistently outperforms abstract descriptions for style matching.
Tip: Run this analysis on both Claude and ChatGPT. Compare the two voice analyses. Where they agree, you've found a real pattern. Where they disagree, look more closely at your samples to see which interpretation is more accurate.
Take Claude's analysis and structure it into a Context Card with these sections.
Tip: Keep the total Context Card under 600 words. Beyond that, the signal-to-noise ratio drops and AI starts treating some rules as more important than others based on proximity to the end of the document.
Load your Voice Context Card into Claude (paste it at the beginning of a conversation or add it to a Project) and generate three different content types: (1) a listing description for a property you've recently sold, (2) a follow-up email to a new lead, and (3) a social media caption for a market update. Compare each output to what you'd actually write. The critical question isn't 'Is this good writing?' It's 'Would I actually say this?' Read the output out loud. If it sounds like you talking to a client, the Context Card is working. If it sounds like a well-written version of someone else, the card needs adjustment. Score each output on a 1-5 scale: 1 = sounds nothing like me, 3 = close but generic in spots, 5 = I'd post this as-is. According to OpenAI's prompt engineering best practices, few-shot examples in system prompts produce more consistent style matching than instruction-only approaches. If your scores average below 3, the Context Card needs more specific constraints or better examples.
Tip: The listing description is usually the hardest test. If your Context Card can produce a listing description that sounds like you, it'll handle emails and social posts easily.
Look at where the AI output diverged from your voice and make targeted adjustments. Common failure patterns and their fixes: Too formal: Add a rule like 'Write at an 8th-grade reading level. Use contractions. Address the reader as you.' Too many adjectives: Add to your constraints: 'Maximum 2 adjectives per sentence. Replace adjectives with specific data when possible.' Wrong structure: Add or revise a few-shot example that demonstrates the structure you want. Generic openings: Add specific opening patterns: 'Start with a market fact, a neighborhood observation, or a client scenario. Never start with a question or a greeting.' Each adjustment should be specific and testable. Don't add vague instructions like 'be more casual'—add concrete rules like 'use contractions always' and 'sentence max: 20 words.' The OODA Loop applies here: observe the output gaps, orient against your actual writing samples, decide on specific rule changes, act by updating the Context Card.
Tip: Keep a running log of every adjustment you make and why. After 3-4 rounds of iteration, you'll have a Context Card that captures your voice with 90-95% accuracy—which is the quality threshold where AI output becomes genuinely useful.
Your core Context Card handles about 80% of voice matching. The remaining 20% is content-type-specific. Create short extensions (50-100 words each) for your most common content types: listing descriptions, client emails, social captions, and market updates. Each extension adds format-specific rules on top of your core voice. For example, a listing description extension might add: 'Lead with the neighborhood and lifestyle, not the property features. Include one hyperlocal detail in the first two sentences. End with square footage and price, not a call to action.' An email extension might add: 'Open with their name. Keep to 3 paragraphs max. Always end with a specific next step and a timeframe.' Store the core Context Card and extensions in a Google Doc, Notion page, or your CRM's template section—somewhere accessible from any device. Load the core card plus the relevant extension before each AI task. This is the system that turns AI from a one-off experiment into a consistent content engine that scales your voice across every channel.
Tip: Update your few-shot examples monthly. As AI produces good output using your Context Card, save the best outputs and rotate them into the card as new examples. Your Context Card gets better over time as the examples get more refined.
I want to build my AI brand voice. Here are 5 writing samples that represent how I actually communicate with clients and on social media: [SAMPLE 1 - Client email after first showing] [paste] [SAMPLE 2 - Instagram caption for a just-sold post] [paste] [SAMPLE 3 - Listing description I wrote] [paste] [SAMPLE 4 - Email to my sphere about market conditions] [paste] [SAMPLE 5 - Text message to a repeat client] [paste] Analyze these samples and create a Voice Context Card with these sections: 1. Voice Identity (2-3 sentences) 2. Voice Rules (5-8 specific, testable rules) 3. Vocabulary Constraints (USE / NEVER USE lists) 4. Structural Patterns 5. Few-Shot Examples (select the 2 best samples to include) Then generate a test listing description using the Voice Context Card for this property: 4BR/3BA in Brentwood, TN. 2,800 sq ft, renovated kitchen, half-acre lot, Williamson County schools. Listed at $875,000.
VOICE CONTEXT CARD — [Your Name], Nashville Real Estate 1. VOICE IDENTITY I communicate like a knowledgeable neighbor who happens to sell real estate. Direct, specific, and grounded in data. I'm warm but never cheesy. I explain market dynamics the way I'd explain them to a friend at dinner—clearly, honestly, and with actual numbers. 2. VOICE RULES - Average sentence length: 10-16 words - Use contractions always (I'm, you'll, that's, it's) - One exclamation mark per piece maximum - Lead with facts or observations, never questions - Use specific numbers: '$450K median' not 'affordable' - Paragraphs: 2-3 sentences max - Address the reader directly (you, your) - Humor is dry and occasional, never forced 3. VOCABULARY CONSTRAINTS USE: specific street/neighborhood names, price per square foot, days on market, year-over-year comparisons, 'here's the thing,' 'the reality is,' 'what that means for you' NEVER USE: stunning, nestled, boasts, dream home, won't last long, act fast, gorgeous, breathtaking, charming, turnkey, entertainer's 4. STRUCTURAL PATTERNS - Open with a market fact or neighborhood observation - Body: alternate between data and interpretation - Close with a specific next step or takeaway - Never close with 'let me know if you have questions' 5. FEW-SHOT EXAMPLES [Best 2 of your provided samples pasted here] --- TEST OUTPUT: Brentwood Listing Description Brentwood's 37027 zip code averaged $412/sq ft last quarter. This one comes in at $312/sq ft—and the kitchen was just gutted and rebuilt. 4 bedrooms, 3 baths, 2,800 square feet on a half-acre lot in Williamson County. The school district is the reason half of Nashville's relocating families land here—Scales Elementary, Brentwood Middle, Ravenwood High. The kitchen renovation is the headline: new cabinetry, quartz counters, and a layout that actually makes sense for a family of four. The half-acre lot is increasingly rare in Brentwood proper—most new construction sits on a quarter acre. $875,000. That's Brentwood with a renovated kitchen and room to breathe. The numbers work.
Your Voice Context Card should produce output you'd rate 4-5 on the 'would I actually say this?' scale for at least 2 out of 3 content types. If you're consistently at 3 or below, your few-shot examples need to be stronger.
The fastest way to refine your Context Card: generate AI output, edit it to sound exactly like you, then add the edited version as a new few-shot example. Each cycle tightens the voice match.
If you work on a team, each agent should have their own Voice Context Card. Your team brand guidelines can be a shared extension that sits on top of individual voice cards—ensuring brand consistency while preserving individual personality.
Claude Projects are ideal for storing your Context Card permanently. Create a Project called 'My Brand Voice' and add the Context Card to the Project Knowledge. Every conversation in that Project will automatically use your voice. ChatGPT Custom Instructions work the same way but have a tighter character limit.
Describing your voice in vague terms: 'I'm professional but approachable'
Fix: Replace vague descriptions with specific, testable rules. Instead of 'professional but approachable,' write: 'Use contractions. 12-word average sentence length. Never use jargon without explaining it. Address the reader as you.' Specific rules produce specific output.
Using marketing copy as writing samples instead of authentic communication
Fix: Marketing copy written by someone else—your brokerage, a marketing agency, or a template—doesn't capture your voice. Use emails you wrote yourself, social posts you typed out, and client messages that reflect how you actually communicate.
Building the Context Card once and expecting it to work forever
Fix: Your voice evolves. Review and update your Context Card monthly. Swap in better few-shot examples as AI produces output you're proud of. Remove rules that are no longer relevant. A voice card should be a living document that improves with use.
Learn the Frameworks
Deep dive into Context Cards—the structured documents that make AI output sound like you.
How to use few-shot examples and structured constraints to match any writing style.
Why showing AI examples of your writing is more effective than describing your voice.
How AI is changing brand voice consistency and what real estate professionals need to know.
Stop guessing with AI. Join The Architect workshop to master the frameworks behind every guide on this site.