The short answer: AI writing sounds robotic because it falls back on safe, average phrases that nobody actually says. To make it sound like you, give it samples of your own writing, write down the rules of your voice, keep both in a Claude Project, and ask for drafts in that voice. Then cut the giveaway phrases and read the draft aloud before you send it.
Most people who use ChatGPT or Claude at work know the feeling. The draft is correct, polite and well organised, and it sounds like nobody you know. Your client notices too. A message that sounds like a machine wrote it gets less trust and fewer replies.
The fix is not a cleverer tool. It is giving the AI what it does not have: your words, your habits and your opinions. This guide shows the house-voice method we use in our private training for teams in Nairobi, and the prompts that go with it.
What are the signs that a text was written by AI?#
AI text gives itself away through a small set of habits: filler openings, a handful of overused words, lists of three, and the same tidy rhythm in every sentence. One habit on its own proves nothing. Five in one short message tell the reader that nobody edited it.
Here are the ten we point out in class, each with a plain alternative.
- 1. Filler openings: "In today's fast-paced world" or "I hope this email finds you well". Start with the point instead.
- 2. Favourite AI words: delve, leverage, seamless, robust, elevate, unlock, tapestry, landscape. Use the word you would say: look at, use, smooth, strong.
- 3. Everything in threes: "fast, reliable and affordable". Give one fact the reader can check: "ready by Thursday".
- 4. "Not just X, it's Y": "It's not just a salon, it's an experience." Say what it is and what it costs.
- 5. Long dashes everywhere: AI uses the long dash far more than most Kenyan writers do. Use a full stop or a comma.
- 6. Headings and bold text in a short message: a WhatsApp reply does not need a heading. Write it the way you would say it.
- 7. Stiff politeness: "Please do not hesitate to reach out." Write "Call me on this number" or "Reply here and I will sort it."
- 8. Empty praise: "Great question!" or "What a wonderful idea!" Answer the question.
- 9. The same rhythm in every sentence: medium length, medium length, medium length. Mix one short sentence with a longer one.
- 10. A summary nobody asked for: "In conclusion" or "Overall". End with the next step and stop.
Why is a humaniser tool the wrong fix?#
A humaniser rewrites AI text to fool an AI detector. It swaps words and shuffles sentences, but it cannot add what makes writing yours: your facts, your opinions and the way you talk to your customers. The result often reads worse, and it can change the meaning.
Humanisers exist because people fear AI detectors. Those detectors are not reliable. OpenAI withdrew its own AI text classifier in July 2023 because of its low rate of accuracy. Its announcement page says the tool caught only 26% of AI-written text and wrongly flagged human text 9% of the time (checked October 2026).
The problem is worse for people who write English as a second language, which is most of us in Kenya. A Stanford study found that seven popular detectors were near perfect on essays by American schoolchildren, but flagged 61% of essays by non-native English speakers as AI-generated. Read the Stanford HAI summary (checked October 2026).
So the goal is not to pass a detector. The goal is a message your client trusts. That needs your voice, not a disguise.
What is the house-voice method?#
The house-voice method teaches the AI how you write before it writes for you. It has three parts: samples of your own writing, a short list of rules drawn from them, and a Claude Project that keeps both, so every new draft starts in your voice instead of the AI's default.
We use it in private training for teams that write to clients every day, and it is also a block of AI Essentials, our one-Saturday AI course in Nairobi. It takes about 30 minutes to set up.
Step 1: Collect five to ten samples#
Pick messages you wrote yourself and were happy with: two client emails, three WhatsApp replies, one proposal paragraph, one social post. Mix formal and casual. Remove names, phone numbers and amounts first, and replace them with labels such as "Client A".
Step 2: Turn the samples into rules#
Paste the samples into Claude and ask it to describe your voice. Then edit its list until it sounds right to you.
Below are seven messages I wrote to clients. Study them and describe my writing
voice as a list of rules another writer could follow. Cover: sentence length,
how I greet and sign off, words and phrases I use often, words I never use,
how formal I am, how I mix English and Kiswahili, and how I ask for the next
step. Quote short examples from my messages for each rule.
[paste your samples]A good rule list is short and specific. "Friendly but professional" is useless. "I greet clients by first name, I never write 'Dear Sir', and I end with one clear question" is useful.
Step 3: Keep it in a Claude Project#
A Project is a workspace in Claude with its own instructions and files. Put your rule list in the instructions and upload your samples as files. Every chat inside that Project now starts with your voice. If you have not built one before, our guide on how to build your own AI assistant walks through it step by step.
You write drafts for me in my own voice. Follow the voice rules below and
match the samples in this Project's files. Never use these words: delve,
leverage, seamless, robust, elevate. No long dashes. No headings or bold in
messages under 150 words. Use British spelling. If you need a fact you do
not have, write [CHECK] instead of guessing.
[paste your voice rules]A worked example#
Riziki Beauty Lounge is an invented salon we use for demonstrations. The owner needs to tell regular customers that braiding prices are going up. Here is the default AI draft, then the draft from a Project that holds her voice rules.
Dear Valued Customer, We hope this message finds you well. We wanted to reach
out to inform you that, in order to continue delivering a seamless and
elevated experience, our braiding prices will be adjusted effective 1st
December. We truly appreciate your understanding and continued support!Hi Wanjiru, a quick heads-up before you book. From 1 December, knotless braids
go up to [new price] because the cost of hair has risen. Book before the end
of November and you pay the old price. Shall I hold Saturday morning for you?The second one is shorter, names a real reason and ends with a question the customer can answer. That is what a voice does.
Which prompts ask the AI for your voice?#
Once your Project is set up, three short prompts do most of the work. Use them inside the Project, or paste your rules into a normal chat if you do not have one yet.
Rewrite this draft in my voice. Keep every fact. Cut anything a busy client
would skip. Make it at least a third shorter.
[paste draft]Check this text for signs that it was written by AI: filler openings, overused
words, lists of three, long dashes, stiff politeness and empty summaries. Show
me each one and suggest a plainer version. Do not rewrite the whole text.Give me two versions of this reply: one shorter and warmer, one more formal
for a new corporate client. Both must sound like the samples in this Project.The second prompt is the most useful habit to build. It keeps you in charge of the edit instead of handing the whole message to the machine. For more prompt patterns, see our guide on how to write a prompt, and our ten AI prompts for small business owners you can copy today.
Why should you read it aloud before you send it?#
Reading aloud is the fastest test of voice there is. If you would never say a sentence to the client's face, rewrite it. Your ear catches stiff phrases, wrong facts and missing warmth that your eye skips over on a screen.
Before you send anything an AI helped you write, ask four questions:
- Would I say this out loud? If not, change the words until you would.
- Is every fact true? Prices, dates, names and promises. AI fills gaps with confident guesses.
- Is there one clear next step? A question, a date, a number to call.
- Is it short enough? Most client messages lose nothing when you cut a third.
This takes one minute. It is also the step that protects you, because the message goes out under your name, not the AI's.
Does this work in Kiswahili and Sheng?#
Yes for Kiswahili, with care. Claude and ChatGPT write good everyday Kiswahili, but it often sounds formal and textbook, not the way people text. Sheng is weaker still: AI tends to invent words or use phrases that are years out of date.
Three habits help:
- Give it your mix. Include samples where you switch between English and Kiswahili, and add a rule such as "greet in Kiswahili, explain prices in English".
- Do not ask it to invent Sheng. If your customers expect Sheng, write those lines yourself, or give the AI phrases you already use.
- Check with a fluent speaker before anything important goes out, especially adverts and public posts.
Frequently asked questions#
Can clients tell that AI wrote something?#
Often, yes, when the text is unedited. Filler openings, overused words and stiff politeness are easy to spot once you know them. When the draft starts from your own samples and rules, and you edit and read it aloud before sending, clients see your message. What they notice is the effort, or the lack of it.
Do AI detectors work?#
Not reliably. OpenAI withdrew its own detector in 2023 because of low accuracy, and a Stanford study found popular detectors flagged most essays by non-native English writers as AI-generated. Do not use a detector to judge a colleague's or a student's work, and do not use a humaniser to beat one.
Is it dishonest to write with AI?#
Not in itself. Using AI to draft is like using a spell checker or asking a colleague for a first version. It becomes dishonest when you send facts you have not checked or claim work you did not do. Follow your employer's or client's rules, and disclose AI use where they require it.
Cloudwise Limited, an ICT company in Westlands, Nairobi, runs hands-on AI training on Claude for individuals and teams.
Leave with your own AI assistant: AI Essentials, one Saturday, Ksh 5,500. You build a Claude Project for the job that takes most of your week, and that includes writing in your own voice. The next class is Saturday 5 December, 9:00 am to 1:00 pm. Book at cloudwise.co.ke/ai-training/essentials.
Written by
Gavin



Comments
Be the first to comment.
Leave a comment