You asked an assistant for a quick email, and it came back stiff, padded and nothing like you. Getting AI to write like you comes down to a short saved prompt plus three real examples of your own writing, and you can set it up without any code. Save it once, and every new draft starts closer to your voice.
The prompt names who you are, who reads your messages, how you sound, which words you never use and what the finished piece should look like. The examples show the assistant your habits at work. The method is the same in ChatGPT, Claude and Gemini, and it takes about ten minutes.
A saved prompt will not make a draft final. It gets you a better first version, and you still read it before a customer does. That is the trade: less rewriting, with your name still standing behind every word you send.
Key Takeaways
Save a short prompt that covers the writer, the reader, the voice rules, the words to avoid, the accuracy rules and the format, then attach three real pieces of your writing.
Examples show the assistant what you sound like, and the rules tell it which of those habits matter, so the result approximates your voice and does not copy it.
Run the prompt on three different jobs, such as an email, a public post and a customer reply, and change one rule at a time.
Never paste passwords, payment details, medical information, private customer records or confidential contracts. Write [CUSTOMER NAME] and [PRICE], then add verified details after drafting.
A saved prompt is reusable instructions, not permanent training. A person checks names, dates, prices, promises and whether it sounds like you.
Spend ten minutes once and stop rewriting every draft
Start with your own writing, because that is where your voice already lives. Open three emails or posts you are proud of, written for a similar audience and purpose, and paste them into one document. Choose them for consistency, not only for polish. Three pieces that sound alike teach an assistant more than three impressive ones that do not.
Next, mark what they share. How quickly do they get to the point? Are the sentences short or long? Is the tone warm, blunt, formal or conversational? Which words keep showing up, and how do you handle greetings, questions and closing lines? Do you use bullets, headings or plain paragraphs?
Turn each repeat into a rule you could hand a new employee. For example: start with the answer, use plain language, avoid exclamation marks, keep paragraphs under three sentences, end with one clear next step. Rules like these work better than a word such as "professional" or "human," because an assistant can follow a rule you can observe.
Copy a prompt template that works in any chat assistant
A useful saved prompt has eight parts. Each one answers a question the assistant would otherwise guess at, and guessing is where the generic tone comes from.
| Part | What you write |
|---|---|
| Writer | Who you are, what the business does and how you normally communicate |
| Reader | Who will read it, what they already know and what they need |
| Voice rules | Tone, sentence length, paragraph length and level of formality |
| Structure | A direct opening, short paragraphs and a clear next step |
| Examples | Three real pieces of your own writing |
| Avoid list | Words, phrases, clichés, punctuation habits and claims you never make |
| Accuracy rules | Do not invent facts, prices, names, guarantees or customer details |
| Output format | Email, post or reply, and how long |
Copy this template into a plain document, then replace every bracket with your own words. The first lines hold what stays the same from job to job. The last block, the task, is the only part you change each time.
```text You are writing for [business and audience].
Writer identity: [Who I am, what I know, and how I normally communicate.]
Audience: [Who will read this, what they know, and what they need.]
Voice rules:
- Use [plain / warm / direct / formal] language.
- Prefer [short / varied] sentences.
- Lead with the answer or useful action.
- Use [paragraph, list, heading] format.
- Explain technical terms in ordinary language.
- Sound [three to five precise adjectives].
Avoid: [Words, clichés, exaggerations, punctuation habits, claims, and tones to avoid.]
Accuracy: Do not invent facts, figures, customer details, prices, sources, guarantees, or legal conclusions. Mark uncertainty and ask for missing facts.
Examples of my writing: Example 1: [paste real email or post] Example 2: [paste real email or post] Example 3: [paste real email or post]
Task: Write [format] about [subject] for [audience]. Apply the voice rules, but preserve the facts and purpose of the task. ```
The accuracy block matters as much as the voice rules. It tells the assistant to ask for what it does not know, so you stop finding invented prices and made-up promises in a draft you were about to send.
Save it where your assistant keeps standing instructions
Each assistant has a place for instructions you want applied again and again. The names and menus change as the products change, so open the assistant's settings or help page and search for the feature name before you follow any older screenshot. Keep the original prompt in a plain document of your own as well, so you can paste it anywhere.
ChatGPT
OpenAI says Custom Instructions let you give ChatGPT information to consider, apply it across chats, and edit or remove it later. Projects can hold project instructions and files, depending on your plan and workspace. Memory is a separate feature from both, described on OpenAI's Memory page.
Claude
Anthropic describes Projects as a place for project instructions and a knowledge base, and those instructions apply to chats inside the project. Claude also has styles, which control how it formats and delivers responses. Anthropic's page on Claude's personalization features explains how the pieces fit together.
Gemini
Google says Gems in Gemini Apps are saved custom versions of Gemini, set up with reusable instructions for goals you repeat. Google's guide to getting started with Gems walks through creating one. A Gem is a good home for your voice prompt and your three examples.
Pick one assistant and start there. If you use two, paste the same document into both so your voice does not drift between them.
Keep customer details out of the prompt and still get usable drafts
A saved prompt travels to a company's servers, so write it as if a stranger might read it. Do not paste passwords, API keys, payment-card details, authentication codes, private customer records, medical information, complete client files or confidential contracts. Leave out any identifying detail the writing task does not need.
Each vendor handles saved content under its own terms. OpenAI says personal ChatGPT content may be used to improve its models depending on your account settings, and that business, enterprise, education and API data are treated differently by default, as its page on how it handles data in consumer services explains. Google says Gemini activity may be stored and used according to your account settings, per its Gemini Apps Privacy Hub.
For Claude, treat anything you attach to a project as business data you are sending to Anthropic under your account terms, as its page on what projects are describes.
The fix costs almost nothing. Write placeholders such as [CUSTOMER NAME], [ORDER NUMBER] and [PRICE] in the draft request, then type the verified details in yourself after the assistant has written the shape of the message. Check your account settings for each assistant before you paste anything you would not want stored.
Trust the first draft by testing the prompt on three different jobs
Once the prompt is saved, test it on three different jobs: one email, one public post and one customer reply. Put each result next to your real writing and look for the same mismatch showing up more than once. A repeated mismatch is a rule problem, and a one-off oddity is just a draft.
Change one rule at a time, then run the same job again. If you change five things at once, you will not know which one helped. Add a new example only when it shows the voice you want now, not merely because it is your most recent message. When your voice or your business changes, swap the examples out.
The voice rules stay the same for every job, and the task block changes. Tell the assistant the audience, the goal, the facts, the length and the call to action each time, and keep the stable rules separate from the job-specific ones.
| Job | What to add to the task block | What to check before sending |
|---|---|---|
| Customer email or quote | The customer's request, the scope, the verified price and the next step | Names, dates, scope and every number |
| Public post | The point of the post, the platform and the length | Factual claims and the platform's format |
| Review reply | The reviewer's actual situation and what you can offer | That it matches what really happened |
A review reply needs the most care, because it has to reflect the real customer situation. Our post on using AI to reply to reviews safely covers that job in detail.
Know what a saved prompt can and cannot do for your writing
Examples give the assistant evidence about your patterns. Your rules tell it which patterns to keep. Together they get you close, and results still vary by task, model, settings and the quality of your examples. Expect an approximation of your voice, not an exact copy.
A saved prompt is also not training. It is a set of instructions the assistant reads each time, so the model itself does not permanently learn from it. It is not a guarantee of authorship, and it does not replace editing. You supply the facts for each job, and you approve the final wording.
A person should check names, dates, prices, promises, legal or regulated claims, customer facts, and whether the wording sounds like you. NIST's generative AI risk profile, AI 600-1 from July 2024, lists data privacy, confabulation (the model stating invented details as fact), information integrity and human oversight among the risks to manage, and it recommends privacy steps such as anonymizing data. If you ever describe your AI use in advertising, the FTC's February 2023 note, Keep your AI claims in check, warns businesses against unsupported claims about what AI can do.
This method is also separate from building a custom GPT, which is a different project described in our post on a custom GPT for your business. If you are still deciding which plan to start on, whether free ChatGPT is enough is a separate question worth reading. And for longer pieces, our post on using AI to write your blog covers what to hand off and what to keep.
Does your saved prompt hold up?
Pick an answer to begin.
1. What belongs in a saved voice prompt?
2. A draft keeps using a phrase you would never say. What should you do?
3. Does saving a prompt permanently train the assistant to write exactly like you?
Frequently Asked Questions About Making AI Write Like You
Can AI write exactly like me?
It can approximate the patterns in your examples and rules, but exact imitation is not guaranteed. Plan to read and adjust every draft.
How many examples should I give it?
Start with three representative pieces. Add more only when they clarify a pattern the assistant keeps missing. Consistency matters more than the exact number.
Do I need a custom GPT to make AI write like me?
No. A saved prompt in the assistant's ordinary instruction features is enough to start, with no code and no builder.
Can I use one saved prompt for every job?
Keep one voice foundation, then add a task block for each job. Emails, posts, quotes and replies each need their own facts, length and call to action.
Should I paste customer emails into the prompt?
Not with names, order numbers, prices or other private details. Use placeholders, and add the verified details after the draft is written.
How do I make AI writing sound less robotic?
Replace vague words like "professional" with rules you can observe, such as short sentences, the answer first and one clear next step. Add three real examples, then test and adjust one rule at a time.
What This Means for You
You can teach an assistant your voice without code. Save one short prompt that covers the writer, the reader, the voice rules, the words to avoid, the accuracy rules and the format, attach three real examples, and test it on three different jobs. Examples show the patterns, rules say which to keep, and the result approximates your voice without copying it.
The payoff is a first draft that starts closer to you, which leaves your evenings for the parts only you can do: checking the facts, the promises and the tone. Keep the prompt current by swapping examples when your voice changes and fixing one rule at a time.
You can write the first version yourself this week. If several people write for your business, or you want shared standards, privacy rules and a repeatable review process, Web Leveling can help through our AI consulting, and our content marketing work covers the writing that follows. We will not promise that an assistant sounds exactly like you, because no one can. We work with small and medium businesses across the country and overseas. Tell us who writes for your business and what must stay private, and we will help you set it up.
Terms
Saved-prompt words in this post
Tap a term to see what it means.
Saved prompt. A short set of instructions you store in an assistant so it applies them every time you ask for a draft.
Voice rules. Observable habits, such as short sentences or the answer first, that describe how you write.
Example. A real email or post of yours that shows the assistant your patterns in practice.
Placeholder. A bracketed stand-in, such as [PRICE], that you replace with a verified detail after the draft is written.
Confabulation. When an AI model states invented details as if they were facts.
Project. A feature in ChatGPT and Claude that holds standing instructions and files for related chats.
