Tagline (gold-shimmer treatment, DM Serif Display 56px, sits above H1):
Learn, build, accelerate.
You can't describe your voice. Show it.
Most REALTORs describe their voice as "professional but warm." That's a vibe, not a Voice Card layer. The model can't write to a vibe.
Show the model. Paste three to five of your past emails, run the Style Match prompt, and the model returns a structured Voice card you can lock into your Context Card. From then on, AI output sounds like you. Not like ChatGPT default. Not like the internet's average.
Six minutes. $20 a month. The fix for the "sounds like a robot" problem we hear in every workshop.
The problem
The 12-deal Cool Springs agent finishes a listing prep on Old Hickory Lake. She types into ChatGPT — write a launch email for this listing, professional but warm. The output is 240 words, two exclamation marks, the word "nestled" twice, and the phrase "luxury living awaits." She deletes the whole thing.
She tries again — write it in my voice. The model has no idea what her voice is. It defaults to a generic-warm-agent template that sounds like every other launch email in the inbox.
That's the bottleneck. Not the model, not the prompt, not the listing. The Voice layer is missing. Every AI output downstream is a fill-in-the-blank against a placeholder named "her voice," and the placeholder defaults to the average of every email the model ever read.
The fix isn't a better adjective. "Professional but warm" doesn't help the model. Neither does "casual but expert." Adjectives are a vibe. The model wants patterns — sentence length, contraction policy, em-dash usage, banned words, hook moves, closing moves.
You have those patterns already. They live in the emails you already wrote.
The Style Match technique
Three steps inside the head, written out for the page.
One — paste the samples. Three to five of your strongest past emails. The model reads the prose.
Two — extract the patterns. The model returns a structured Voice card — sentence length, contractions, em-dash policy, banned words, hook patterns, opening and closing moves.
Three — edit and lock. You read the output, edit two or three lines that don't match how you actually write, and paste the result into the Voice layer of your Context Card.
That's the whole technique. The output is a working Voice card you keep — paste it into every future AI session, and the output stops sounding like ChatGPT default and starts sounding like you.
This is what we run live in Section 5 of the AI Acceleration class. Same exercise on the workshop circuit. Same exercise inside the Listing Machine cohort. The technique is durable. The model behind it is whichever foundation model you're paying for this quarter.
The prompt
Copy this verbatim. Don't edit the constraints.
You are a writing-voice analyst. I'm going to paste 3-5 of my
past emails. Read them and extract a structured Voice card I can
load into a Context Card for AI tools.
Output the Voice card in this exact format:
VOICE CARD — [my name or initials]
Channel: [email / listing copy / social / spoken — pick from the
samples]
Sentence length pattern: [short examples — average word count,
range, what triggers a long sentence]
Contractions: [which ones, how often, exceptions]
Em-dash policy: [used / not used / how often, what for]
Punctuation rules: [exclamation marks count, semicolons used,
ellipses, parentheses]
Hook patterns: [how the writer opens — examples from the samples]
Closing patterns: [how the writer closes — examples from the
samples]
Banned words / phrases (do not say): [words the writer never uses
or specifically avoids — list 8-15 if you can find them]
Concrete proper nouns the writer uses: [names of neighborhoods,
streets, schools, brokerages, MLS abbreviations — these are the
local-knowledge anchors]
Voice essence in one sentence: [a one-sentence summary the writer
would actually agree with]
Two example sentences in this voice (write new ones, don't quote
the samples): [show me you can produce in this voice]
Rules:
- Quote specific examples from the samples to back every claim.
No vague adjectives like "professional but warm."
- The banned-word list is the most important section. If a word
doesn't appear in any of the samples, that's signal — list it
as banned.
- Do not editorialize about the writer's personality.
- Do not add emojis to the output.
- Output as plain text I can paste into a markdown file.
After the prompt, paste your three to five emails. Each separated by a divider line — three dashes on a blank line is enough. Strip the signature blocks and the email-client headers. The model wants prose.
Walkthrough — Brentwood listing-launch emails
Real example. Working agent in Brentwood, 14 deals last year, sphere-driven. We ran the Style Match against three of her listing-launch emails — one for a Westhaven bungalow, one for a Concord Park new-build, one for a Old Hickory waterfront across the county line. Six minutes, end to end.
Step 1 — She picked the three emails
Sent folder. Filter by "subject contains 'just listed' or 'launching.'" Top of the list. She picked the three she remembered writing on a good day — not the rushed ones, not the ones the assistant drafted. The Westhaven one ran 210 words. The Concord Park one ran 180. The Old Hickory one ran 240.
The pattern test — read all three out loud. They sound like the same person. If they don't, you're picking from too many channels and the extraction will average to mush.
Step 2 — She pasted the Style Match prompt
Fresh ChatGPT chat. Pasted the prompt above verbatim. No edits.
Step 3 — She pasted the three emails after the prompt
Three dashes between each. Stripped the "Sent from my iPhone" footer and the brokerage email signature. Just the body copy.
Step 4 — ChatGPT returned the Voice card in 90 seconds
The output looked like this. (Names redacted, structure intact.)
VOICE CARD — [agent initials]
Channel: listing-launch email
Sentence length pattern: average 14 words. Range 6-28. Long
sentences only when listing 3+ specific features in one breath
(e.g., "Three bedrooms, two baths, original 1996 oak floors,
covered back porch, dock with electric.")
Contractions: always. "It's", "we're", "won't", "she'd". Never
"it is" / "we are" / "she would" except in formal closings.
Em-dash policy: used. About one per 80 words. Almost always for
a parenthetical aside ("the kitchen — fully redone in 2023 — is
the room buyers tour twice").
Punctuation rules: zero exclamation marks across the three
samples. No semicolons. No ellipses. Parentheses used sparingly
for asides about the neighborhood.
Hook patterns: opens with the listing's specific anchor.
Examples — "The Westhaven listing is live." "5847 Concord Park
just hit MLS." "Old Hickory waterfront, dock included, pre-list
walkthrough Saturday." Never opens with "I'm so excited" or
"happy to share."
Closing patterns: ends with a specific next step and a name.
Examples — "Saturday 10-11. Park on the right side of the
driveway. Text me. — [name]" "Walkthrough this weekend, comp
sheet attached. — [name]" Never closes with "let me know if
you have any questions."
Banned words / phrases (do not say):
- nestled
- welcome home
- luxury living awaits
- truly stunning
- a must-see
- don't miss out
- act fast
- your dream home
- entertainer's delight
- "I'm thrilled"
- "I'm so excited"
- exclamation marks (zero in samples)
- emojis (zero in samples)
Concrete proper nouns the writer uses: Westhaven, Concord Park,
Old Hickory Lake, Brentwood, Cool Springs, Williamson County,
MTSU, MLS, Compass
Voice essence in one sentence: an operator briefing a buyer's
agent on what's live, what matters, and what to do next — no
adjectives, all specifics.
Two example sentences in this voice:
1. "The Cool Springs relisting goes live Tuesday — comps are in
line, the kitchen redo is the move, walkthrough Wednesday."
2. "Brentwood split-level on Hill Road, three offers in the
first 48 hours, we'll know by Friday."
That's the output. Read it once. The agent edited two lines — the model claimed she used parentheses "sparingly," but she actually doesn't use them at all. She corrected it. The model claimed her closing always names herself — true for two of three samples, but the third closed with just initials. She loosened the rule.
Five minutes total. Voice card locked.
Step 5 — She loaded it into her Context Card
The Voice card slots into the Voice layer of the 4-layer Context Card — Role, Voice, Do-Not-Say, Local Knowledge. The Style Match output produces three of those four layers in one pass. The Role layer she already had on a clipboard. The Local Knowledge layer she'd built from a previous session. Now the four layers stacked. The full Context Card pastes into any AI tool — ChatGPT, Claude, Gemini — at the start of every working session.
Next time she briefs AI for a listing email, the output ships in 30 seconds of editing instead of 5 minutes of fighting the default.
What the output looks like in practice
The Voice card above replaces "professional but warm" with mechanical rules.
Before — write a launch email for this listing, professional but warm.
After — paste the Context Card. Then write a launch email for this listing.
Same listing. Different output. The first one ships generic ChatGPT copy. The second ships copy that sounds like the agent on her best day.
That's the whole game. Andrej Karpathy frames this as Software 3.0 — English is the new programming language. The model is the runtime. The brief is the program. The Style Match output is the part of the brief that locks voice fidelity in. Skip it and you're writing software with placeholder strings.
Ethan Mollick at Wharton calls this the briefing mindset. Treat AI like a new hire — the more concrete the upstream brief, the better the downstream output, with diminishing returns past a few thousand words. The Voice card is the densest brief layer per token. Three or four hundred words of Voice card replace tens of thousands of words of failed prompts.
Anthropic's prompt engineering docs recommend the same pattern under different vocabulary — they call it "providing examples," "system prompts," and "tone-and-style instructions." Same shape. The Style Match technique just industrializes the extraction step so you don't have to write the Voice card from scratch.
Simon Willison's writeup on prompt patterns covers the same territory — show the model what good output looks like, don't try to describe it. The samples are the showing. The Style Match prompt is the asking.
How to use the extracted Voice in future prompts
The Voice card is one-fourth of your Context Card. Layer it with the other three.
Role layer — who you are. Years in business, brokerage, specialty area. ("WCAR-affiliated REALTOR, Compass Brentwood, 14 deals/yr, listing-side specialist on Williamson County waterfront.")
Voice layer — the Style Match output above. Patterns, banned words, hook and closing moves.
Do-Not-Say layer — your hard banned-word list. Some of this overlaps with the Voice layer's banned words. Keep both — the Do-Not-Say layer is the absolute floor, the Voice layer banned words are the strong-preference floor.
Local Knowledge layer — proper nouns, market quirks, neighborhood-specific vocabulary. ("Cool Springs traffic on 65 between 4 and 6 PM. Brentwood-zip-but-actually-Franklin properties. The Hendersonville waterfront submarket on Old Hickory Lake. WCAR vs GNAR boundaries.")
The four layers paste into any AI session as the system prompt or the first message. Then you ask for the asset. The output downstream sounds like you, knows your market, avoids your banned words, and uses your hook and closing moves.
This is also the foundation of the foundation-model workflow without Zapier — once the Context Card is loaded, you don't need an automation tool to make AI sound like you. The brief does the work that automation can't.
For the large language model underneath — Claude, ChatGPT, Gemini — the Voice layer is the difference between a model that's writing for the internet's average and a model that's writing for your specific reader.
When to re-extract
Voice drifts. Re-run the Style Match every six months, or any time you change channels.
Six-month refresh. Your writing changes. New listings, new clients, new patterns. Run the prompt against three emails from the last 30 days. Compare against your locked Voice card. If five of the patterns shifted, lock the new card. If only one shifted, edit the live card.
New channel. Email voice and Reels voice are different. Run a separate Style Match for each — three transcribed voice memos for the spoken Voice card, three Instagram captions for the social Voice card. Stack them in your Context Card library. Use whichever matches the channel you're shipping to.
New brokerage. Voice drifts after a brokerage move. The Compass tone is different from the Keller Williams tone, which is different from a regional independent. Run the prompt fresh after a transition. Six months later you'll see the drift in the samples.
After a deliberate voice shift. If you've decided to write punchier, longer-form, more direct, less corporate — whatever the shift — run the prompt against the new samples. The Voice card should match where you're going, not where you've been.
The prompt-engineering glossary entry has the deeper take on iteration patterns. The short version — your Voice card is a working document, not a forever artifact.
FAQ
What if my emails are inconsistent?
Most agents' emails are inconsistent. The closing-day email sounds different from the price-reduction email, which sounds different from the holiday card. That's normal. Pick three to five from one channel and one situation — three listing-launch emails, three follow-up emails, three closing-day notes.
The model will extract the pattern from that channel. You'll build a separate Voice card for each major channel over time. Don't try to extract one universal voice from a mixed bag — the output will be a vague average that doesn't match any of them. Channel-specific Voice cards beat one universal card every time.
If you genuinely have no consistency anywhere — every email is a new tone — that's a different problem. Pick the three you wrote on your best days and treat the output as your aspirational voice, not your average voice. Use the Voice card as a goal, not a description.
How many email samples is enough?
Three is the floor. Five is plenty. More than seven and the model averages too aggressively, which dilutes the strongest patterns.
The samples that count are the ones where you sound the most like you on a good day. One masterpiece email beats four mediocre ones. Pick the three that someone who knows you would read and say "yeah, that's how she writes."
If you've only got two strong samples, write a third in the moment — sit down and draft a launch email for a hypothetical listing in your strongest market. That's a valid sample. The model can't tell the difference between an email you sent last March and an email you wrote this morning.
What about voice memos versus written?
Voice memos give you a second Voice card — the spoken one. Run the same Style Match prompt against three transcribed voice memos and you get a different output.
Spoken voice has more contractions, shorter sentences, more sentence fragments, and a different rhythm. The closing pattern is usually softer. The hook is usually a question. For Reels and YouTube content, the spoken Voice card beats the written one. For email and MLS copy, the written Voice card wins.
Build both. Use whichever matches the channel you're shipping to. The Listing Machine cohort builds three Voice cards by week two — written, spoken, and a hybrid for short-form social where the cadence sits between the two.
Will AI mimic me too well?
The output sounds like you. That's the goal.
The risk people mean when they ask this is impersonation — that someone could clone your voice from public content. The Style Match technique runs on emails you already wrote and own. The output stays in your Context Card, not in the public training set. Your Voice card isn't shipped anywhere. It's a working tool inside your AI sessions.
The bigger risk is the opposite — agents who don't extract their voice keep shipping ChatGPT-default copy that sounds nothing like them, and clients notice. The mimicry risk is low and managed. The dilution risk is high and ignored.
For voice-cloning concerns at the audio layer — that's a different category. See the Section 8 security material for the out-of-band verification and Safe Word protocols.
Can I use this for someone else's voice?
Technically yes. Ethically, only with their consent and inside a working relationship.
Legitimate uses — a transaction coordinator extracting their broker's voice to draft on their behalf. A marketing manager extracting the team lead's voice for the brokerage's social account. An assistant extracting the agent's voice for follow-up emails. In all three, the relationship sets the boundary, not the technology.
What's not legitimate — running the prompt against a competitor's published emails, against a public figure to fake their content, against anyone without their explicit consent. Don't do that. The Style Match technique works the same on any prose — the constraint is the relationship, not the model.
For brokerage-level Voice card systems — where the firm wants a shared voice library across the team — that's the kind of install the Empire program ships. The brokerage owns the master voice and each agent's individual Voice cards layer on top.
What to do this week
Block 15 minutes Tuesday morning. Open the sent folder. Pick three of your strongest emails. Open ChatGPT. Paste the prompt. Drop the emails in. Read the output. Edit two lines. Paste the result into your Context Card.
That's the install. After this week, every AI session you run starts with a Context Card paste. Every output downstream sounds like you. The "sounds like a robot" problem disappears the moment the Voice and Do-Not-Say layers are loaded.
For the wider Context Card system — Role, Voice, Do-Not-Say, Local Knowledge — see the glossary spec. For the workshop where we run this exercise live, see the workshop page. For the apex install where every working layer ships against your real listings, see the Listing Machine.
The Voice card is the layer that pays first. Build it Tuesday.
Sources
Independent creators:
Vendor docs:
Last updated 2026-05-01.
Banned-Word Audit — CLEAN
Audited against the launch-context-card banned list. Confirmed absent from copy: leverage (verb), unlock, unleash, supercharge, empower, elevate, game-changer, revolutionary, transformative, paradigm-shift, best-in-class, seamlessly, effortlessly, robust, synergy, synergize, holistic, ecosystem (as "stuff"), thought leader, thought leadership, industry leader, visionary, "in today's fast-paced market," "in today's digital age," "ever-changing landscape," "dive deep," "deep dive," "dive in," "let's be honest," "to be frank," "look,", "the truth is," "at the end of the day," "when all is said and done," "it's worth noting," "it's important to note," "level up," "next level," "10x," "crushing it," "AI-powered" (as generic adjective), "powered by AI," "the future is now."
Note — banned-word list inside the worked example output ("nestled," "welcome home," "luxury living awaits," etc.) is a literal model output illustrating the Voice card structure. Not used in body copy.
No exclamation marks in body copy. No semicolons. Em-dashes used per spec. Sentences average under 20 words.
Word count (body, excluding front-matter and audits): ~2,150 words.