What to Give ChatGPT So It Sounds Like You
The short answer
Give ChatGPT 8 to 12 recent examples that show how you communicate across different jobs, then turn those examples into a Voice File with evidence-backed rules. Keep changing business facts in a separate, dated Business Brain. Put behavior rules in instructions, reference files in a Project, and private client material nowhere unless your organization has approved the exact tool and data flow.
The input problem
I gave ChatGPT 12 published pieces and nine unedited voice memos.
That was enough material to expose something uncomfortable. My edited writing was punchier than my natural speech, but it had also become too chopped up. The voice memos carried longer sentences, more contractions, and the specific machinery I name when I am trying to make a point land.
I used to think a better prompt would fix generic output. I was wrong about that part. A prompt can assign a task, but it cannot recover language patterns, audience instincts, or business facts it has never been given.
ChatGPT does not usually sound generic because you forgot to type “witty, authentic, and professional.” It sounds generic because those adjectives are doing the job of evidence, and they are spectacularly underqualified.
A Voice File tells ChatGPT how you speak. A dated Business Brain tells it what it may say.
Those are two different jobs. When you mash them into one enormous prompt, durable language rules sit beside an expiring offer, an old service area, and a market fact that was current six months ago. The file grows. Trust drops. Nobody knows which sentence is still true.
The better approach is smaller, separated, and testable.
Separate the jobs
What belongs in a Voice File versus a Business Brain?
The Voice File
This holds communication behavior that should remain recognizable across topics: rhythm, contractions, opening habits, preferred proof, phrases you use, phrases you avoid, humor rules, audience shifts, and known failure patterns.
Example: “Open with a number, a named person, or something I did. Never open with a rhetorical question.”
The Business Brain
This holds current business truth: service areas, audiences, offers, prices, proof, market sources, partner context, frequently asked questions, and claims that need verification. Put a review date on it.
Example: “Monthly engagements begin at $1,500 as of September 2026. Verify before publishing.”
OpenAI makes the same practical distinction in its product guidance. Instructions define how the assistant should behave. Uploaded knowledge gives it source material to use. Project instructions apply only inside that Project and override global custom instructions.[1] [2]
That hierarchy matters when you work in real estate or mortgage. Your voice can remain steady while a listing, lender program, price, market statistic, team role, or disclosure boundary changes. If ChatGPT cannot tell the permanent rule from the temporary fact, neither can the person reviewing its draft.
Collect evidence
How many writing samples should you give ChatGPT?
Start with 8 to 12 short, recent artifacts totaling roughly 1,500 to 3,000 words or transcript equivalent. More is not automatically better. Forty random posts can teach the model your old habits, a contractor's phrasing, and three versions of a brand you stopped using.
Your samples should cover different communication jobs. Include an edited client email, a redacted text message, a voice-memo transcript, a local-market explanation, a service explanation, a published caption, and messages written for different audiences. A Realtor talking to a nervous seller should not sound identical to a loan officer explaining a process to a referral partner.
The mechanism that keeps this clean is a source ledger. It records what each sample is, when it was created, who owns it, whether it was redacted, and what it can teach the model.
| Artifact | Date | Audience | Owner | Privacy | What it teaches |
|---|---|---|---|---|---|
| Seller update email | 2026-08-14 | Anxious seller | Agent | De-identified | Calm sequencing and direct next steps |
| Market video transcript | 2026-08-22 | Triangle homeowners | Agent | Public | Natural sentence rhythm and local language |
| Referral-partner note | 2026-09-01 | Loan officer partner | Agent | Rewritten and approved | Peer tone without corporate filler |
| Service page | 2026-09-05 | Prospective client | Business | Public | Positioning, proof, and offer language |
What should you leave out?
Exclude ghostwritten copy you would never say aloud, stale bios, brokerage boilerplate, captions assembled from trending phrases, and anything you cannot legally or ethically share. A Voice File should teach the model your judgment. It should not become a very organized graveyard for every piece of copy attached to your name.
Use a file gate
Can you upload client emails or mortgage files to create a Voice File?
Do not upload raw client threads, customer relationship management exports, loan files, financial records, identification documents, access details, or property records containing protected information to general ChatGPT. Your Voice File needs language patterns. It does not need another person's private record.
Turning off model training changes how new conversations may be used. It does not give you permission to disclose client information. Temporary Chats are not used for training and are deleted after 30 days, but that still does not replace brokerage policy, lender policy, a contract, consent, or professional judgment.[3]
The Federal Trade Commission's practical rule is useful here: know what information you have, keep only what the job requires, protect what remains, and properly dispose of working copies you no longer need.[4]
| Lane | Examples | Action |
|---|---|---|
| Green | Self-authored public posts, published articles, public videos, approved de-identified examples | Label the source and use it |
| Amber | Client emails, text excerpts, team messages, partner communication, internal training notes | Redact, transform, obtain permission, or get source-owner review before use |
| Red | CRM exports, raw client threads, borrower applications, credit data, financial documents, passwords, access codes, copied competitor work | Do not upload to general AI |
A brokerage AI policy should name approved tools, restricted data, required human review, training, oversight, and an incident path. The National Association of REALTORS® makes the responsibility plain: AI output does not remove the agent's duty to verify accuracy, protect client information, and apply professional judgment.[5]
The implementation sequence
How to create your ChatGPT Voice File
-
Collect 8 to 12 representative artifacts
Aim for 1,500 to 3,000 words or transcript equivalent. Cover several audiences and communication jobs. Choose recent work that you wrote, approved, and still recognize as yours.
-
Redact and label every item
Remove names, contact details, transaction facts, financial details, access information, signatures, hidden spreadsheet content, and unique facts that could identify a person. Add the source-ledger fields before the file enters an approved tool.
-
Run an evidence-first extraction prompt
Do not ask for a flattering personality summary. Require quotations, contradictions, confidence levels, and missing audience coverage.
Analyze the attached approved writing samples to create a Voice File. For every claimed voice trait: 1. Name the trait in plain language. 2. Quote two short examples from different source items. 3. Explain the repeatable writing behavior shown by those examples. 4. Mark contradictions and explain when the voice changes by audience or channel. 5. Label the confidence as high, medium, or low. 6. Flag any audience or communication job that is missing from the samples. Then draft a Voice File with these sections: - Five evidence-backed voice traits - Sentence rhythm and paragraph habits - Vocabulary and phrases I use - Words and phrases I avoid - Opening and closing habits - Humor rules - Audience modes - Jargon translations - Known failure patterns - Final pre-publish checklist Do not infer business facts from writing style. Do not invent a trait because it sounds flattering.
-
Approve the Voice File
Keep five strong traits rather than fifteen vague ones. Add sentence rhythm, actual phrases, banned language, audience modes, humor boundaries, opening habits, closing habits, and the ways your writing goes wrong when it is rushed or over-edited.
Mini Voice File example
Trait: Mechanism before motivation. Explain what failed using a specific object, field, setting, number, or handoff before offering encouragement.
Opening rule: Start with a number, a named person, or something the writer personally did.
Banned language: Avoid empty transformation verbs, vague trend openers, and any claim that could apply to every business.
Failure pattern: When over-edited, the writing becomes a stack of one-line paragraphs and loses its natural sentence rhythm.
-
Create the dated Business Brain
Document service areas, audiences, services, current offers, dated proof, approved market sources, partner context, frequently asked questions, and every claim that requires a fresh check. Add an owner and last-reviewed date.
-
Put each item in the correct place
Put stable behavior rules in Project instructions. Put the Voice File and Business Brain in clear, text-forward reference files. Put the immediate job, audience, current facts, channel, and deadline in the task prompt. OpenAI states that Project instructions override global custom instructions, so duplicated rules can conflict instead of helping.[1]
If you want the larger operating structure around this work, the AI Operating System for Realtors page explains how voice, business context, workflows, channels, and maintenance fit together.
-
Test unfamiliar situations
Do not test by asking ChatGPT to rewrite a sample it just read. That proves it can echo. Test whether the approved files help it handle a new situation without losing facts, tone, or judgment.
Do not grade by vibes
How do you know whether ChatGPT sounds like your business?
Run four unfamiliar-situation tests: a first-time buyer explanation, an anxious seller update, an agent-partner note, and a borrower education message. Give the model a small verified fact block for each task. Do not give it a finished example to copy.
Score every draft from zero to two on five criteria. A zero means it failed. One means it needs meaningful editing. Two means it is accurate and recognizable with only normal review.
A passing setup scores at least eight out of ten on all four scenarios, with no factual or privacy failure. If a draft misses, repair one layer at a time. Change the source when the evidence is weak. Change the Voice File when the pattern is wrong. Change the Business Brain when the facts are stale. Change the instructions when the model ignores a rule.
Add a blind-comparison test
Give one AI draft and one real sample to a person who knows your communication well. Remove the labels. Ask which one sounds more like you and why. The explanation matters more than the guess because it reveals the missing signal: warmth, specificity, restraint, sentence shape, humor, or audience fit.
If every miss gets answered with “try harder,” the setup never improves. You are using ChatGPT like a Magic 8 Ball with a style guide taped to it.
Keep the source current
How often should you update a ChatGPT Voice File?
Review the Voice File quarterly. Update it within one week when your role, audience, approved wording, humor boundaries, or communication style changes in a meaningful way.
Review the Business Brain more often because it holds changing truth. Update it within one week after a change to services, pricing, market coverage, team roles, lender programs, offers, disclosure boundaries, proof, or approved sources.
Add a tiny revision log with four fields: date, owner, changed fact or rule, and test result. The date proves someone reviewed the file. It does not prove every fact inside it remains true.
Quarterly check: Retire stale examples, add one recent artifact for every important audience, check banned language, verify business claims, rerun the four scenarios, and record what changed.
Frequently asked questions
ChatGPT Voice File questions
How many writing samples should I give ChatGPT to learn my real estate voice?
Use 8 to 12 short, recent artifacts that show different communication jobs instead of 40 random posts. Include an edited client email, a redacted text, a voice-memo transcript, a local-market explanation, a service explanation, a published caption, and role-specific messages. Label every item with its date, audience, owner, redaction status, and whether it represents you at your best.
Can I upload client emails or mortgage files to create a Voice File?
Do not upload raw client threads, CRM exports, loan files, financial records, identification documents, access details, or property documents containing protected information. Start with public material or manually de-identified excerpts, remove names and transaction details, and follow the approved-tool policy of the brokerage or lender. The Voice File needs language patterns, not another person's private record.
What belongs in a Voice File versus a Business Brain?
Put durable communication behavior in the Voice File: rhythm, phrases, avoided language, opening and closing habits, humor, audience modes, jargon translations, and failure patterns. Put changing truth in a dated Business Brain: markets, services, current offers, local facts, sourced proof, frequently asked questions, and claims that require current verification. Put behavior and refusal rules in Project or custom instructions.
How do I know whether ChatGPT sounds like my business?
Run four unfamiliar-situation tests and score each draft for factual preservation, audience fit, recognizable language, banned-language avoidance, and the correct next action. Then ask someone who knows the business to compare an unlabeled AI draft with a real sample. A failure means the source, rule, fact placement, or workflow needs repair. It does not mean the model should be told to try harder.
Find the missing layer
Generic output is usually a symptom.
The weak spot may be your Voice File, business context, workflow, channel rules, or maintenance rhythm. The audit shows you where the breakdown starts.
Take the Free AI Readiness Audit