By KEY TAKEAWAYS
- AI sounds generic because it averages everyone; examples of your best writing teach it your voice faster than adjectives like 'friendly but professional'.
- A reliable setup has four parts: a one-page voice doc, a bank of real examples, a banned-phrases list and a second-pass editor prompt.
- Store the setup where the work happens, such as a Claude Project's knowledge and custom instructions, then test it against a fixed set of prompts every time you change it.
Ask five AI tools to write a LinkedIn post for your brand and you will get five posts that could belong to any brand. Same rhythm, same “In today’s fast-paced world” energy, same tidy three-point list. It is not that AI can’t write. It is that nobody told it who you are.
The fix is not a cleverer prompt. It is a small system: a voice doc, an examples bank, a banned list and an editor pass. Build it once, and every draft starts 80% closer to you. (That 80% is our rough working experience, not a measured statistic.)
Here is the exact setup we use, and how to make it stick in the tools your team already has.
Why does AI sound generic in the first place?
Because it predicts the most likely wording, and the most likely wording is average. A model trained on a huge amount of text drifts toward the middle of everything it has seen. If your instructions are also average (“friendly, professional, engaging”), the output is average squared.
Specific input breaks the pattern. Telling the model how you sound, with rules and real samples, gives it something to imitate besides the internet’s default.
What goes into a brand voice system?
Four assets: a voice doc, an examples bank, a banned-phrases list and an editor prompt. Each does a different job.
| Asset | Job | Length | Update |
|---|---|---|---|
| Voice doc | Rules for tone, structure and word choice | 1 page | Quarterly |
| Examples bank | Real writing that shows the voice working | 3-5 per content type | When you write something great |
| Banned phrases | Stops the clichés and off-brand habits | 20-50 items | Whenever you spot a new one |
| Editor prompt | Second pass that checks a draft against the rules | 1 prompt | When rules change |
How do you write a voice doc that works?
Write rules the model can follow, not adjectives it can nod at. “Confident” means nothing. “State an opinion in the first sentence, then give one reason” is something it can do.
Use this template:
- Who we are and who we write for (two sentences).
- Voice in five rules, each testable. Example: “Sentences average under 18 words. We use ‘you’ and ‘we’. We name the risk, not just the benefit.”
- Words we use and words we avoid.
- Formatting habits: how we open, whether we use lists, how we end.
- Hard limits: no invented statistics, no made-up client names or quotes, no medical or legal promises.
If you don’t have a voice doc, ask the AI to draft one from your best writing, then edit it by hand. It will spot patterns you didn’t know you had (and flatter you a little).
Why do examples beat instructions?
Because showing a pattern teaches faster than describing it. Give the model three to five of your best pieces for each content type, plus a one-line note on why each works. This is often called few-shot prompting: a few examples in the prompt that the model copies the pattern from.
Tips from our own use:
- Pick range, not repetition. Include a short post, a long post and an email, not five near-identical captions.
- Add one bad example with a note: “This is what we never sound like.” Negative examples are clarifying.
- Use real writing by the person whose voice you want, not a polished brand-committee version.
How do you build a banned-phrases list?
Read ten AI drafts and highlight everything that made you wince. Those are your entries. Common offenders include openers like “In today’s fast-paced digital landscape,” verbs like “unlock” and “leverage,” and the tidy closing line that restates the whole post.
Add structural bans too: “no rhetorical question in the first line,” “no more than one exclamation mark per piece.” Keep the list in your project instructions and extend it monthly.
Where do you store it so people actually use it?
In a workspace the whole team opens by default. In Claude, Projects are self-contained workspaces with their own knowledge base and custom instructions, so your voice doc can live in the instructions and your examples in the project knowledge (Claude Help Center). Per Anthropic’s documentation, sharing a project with teammates requires a Team or Enterprise plan, so check what yours includes.
Other assistants offer similar features, such as custom instructions or custom assistants. Features and limits change often, so confirm in each vendor’s current documentation before you build around one.
No project feature available? Paste the voice doc and examples into a saved prompt template. It is clunky, but it works, and our Prompt Vault has starter templates.
How do you test whether it worked?
Use a fixed test set and compare before and after. Create five prompts that cover your main content types: a social post, a cold email, a product description, an article intro, a reply to a complaint. Run them after every change to your voice system.
Score each output from 1 to 5 on three questions:
- Would our team believe a colleague wrote this?
- Is any banned phrase present?
- Is there a specific detail, or is it all general?
Keep the scores in a sheet. When quality drifts, you will see it in the numbers rather than in an awkward client email.
What does the editor pass look like?
A second prompt that critiques the draft against your rules and rewrites only what breaks them. Run it after the first draft:
You are our brand editor. Using the voice rules and banned list in this project, review the draft below. List every violation, then return a corrected version that changes as little as possible. Flag any claim or statistic that is not supported by a source in the draft, and do not add new facts.
The last instruction matters. Models happily invent numbers to sound authoritative, so the editor pass doubles as a fact-risk flag. It does not replace a human check of facts.
What does this look like on a real task?
Here is a before-and-after pattern, using an invented example for illustration. Say you run a bookkeeping firm and want a LinkedIn post about quarter-end deadlines.
Without a voice system, the draft opens: “Quarter-end can be a stressful time for businesses.” It is true, bland and interchangeable.
With a voice doc that says “open with the cost of the mistake, use one concrete scenario, never open with a general statement,” the draft opens: “The most expensive spreadsheet error is the one found in April.” Same topic, same model, different instructions. The edit you make after that is small, and that is the goal: your time goes into judgment and a first-hand detail, not rewriting boilerplate.
The same logic applies by channel. Your email voice is probably more direct than your LinkedIn voice, and your support replies are warmer than both. Add a short “channel notes” section to the voice doc, one or two lines per channel, rather than writing separate documents.
How much should you let AI write?
Let it draft structure and first passes; keep the opinions, stories and numbers human. AI is good at turning your notes into clean paragraphs and bad at knowing what you have personally seen. The pieces that earn trust contain something only you could write: a lesson from a project, a pricing decision, a mistake. We add that line by hand to every post, and so should you.
What are the common mistakes?
- Too much material. Fifty mediocre samples teach less than five great ones.
- Voice by committee. If the doc contains every stakeholder’s preference, it describes nobody.
- No human sign-off. AI drafts; a person approves. Our agent guardrails guide covers review steps for anything customer-facing.
- Set and forget. Brands evolve. Review the system each quarter.
For the bigger picture of where this fits, see The Complete Guide to AI Marketing, and if you would rather have this built into your workflows, that is what our AI marketing workflows service does.
What to do this week
- Collect five of your best pieces (a post, an email, an article intro) and write a one-line note on why each works.
- Draft a one-page voice doc with five testable rules, and list 20 phrases you never want to see again.
- Set up a project or saved prompt that includes the doc, the examples and the banned list.
- Create your five-prompt test set and score the first outputs.
- Add the editor prompt as a standard second step, with a human approving anything public.
Frequently asked questions
How do I make AI write in my brand voice?
+
Give it three things: a short voice guide with do and don't rules, three to five real examples of your best writing per content type, and a list of phrases you never use. Put them in a reusable workspace, such as a Claude Project, then test with fixed prompts and refine wherever the output misses.
Can ChatGPT or Claude learn my writing style?
+
Within a chat or project, yes: they imitate the style of the instructions and examples you provide. They don't permanently retrain on your data from your prompts. You need to supply the guide and examples each time, or store them in a project, custom instruction or similar feature, and check each tool's current documentation.
How many writing samples do I need to train AI on my voice?
+
Start with three to five strong examples for each content type, such as emails, social posts and long-form articles. More is not always better; a few excellent, varied samples with a note on why each works usually teach more than a pile of average ones. Add negative examples of what you don't want.
Why does AI writing still sound generic after I add a style guide?
+
Adjective-based guides like 'warm and professional' describe almost every brand, so the model falls back on its defaults. Concrete rules fix this: sentence length, words to avoid, how you open and close, and real before-and-after examples. An editor pass that removes generic phrasing also helps.
Should I publish AI-written content without editing?
+
No. Treat AI as a drafting partner and keep a human approving anything public. Check facts, remove invented claims and add a first-hand detail only you have. A review step also protects your brand from confident errors, and it keeps the final piece sounding like a person with opinions.
Sources
- Anthropic, What are projects? | Claude Help Center
WRITTEN BY
Junaid Mubashar
Founder of an AI-native growth agency. Junaid plans and runs full-funnel marketing (SEO and AI search, paid social, Google Ads) and builds the AI workflows and agents that automate it. Everything on this blog is tested with real campaigns before we recommend it.
🗝️ The Marketer's Prompt Vault
120 battle-tested prompts for Claude & ChatGPT — research, copy, ads, SEO, email and analysis.