Steer the voice of the output using samples of your own writing and examples of what not to do.
Find a reply your AI assistant drafted for you recently and read it next to one you wrote yourself to a similar message. Compare how each one opens, how long the sentences run, how formal the wording is and how it signs off. Those differences are what you want the assistant to stop producing. Lesson 2.1, "One good example beats a paragraph of description", showed that examples set the pattern for structure and content. Voice works on the same principle, with a few extra techniques on top.
Daniel runs a small business and answers customer messages himself. His assistant already tags incoming messages well, so he starts asking it to draft replies too. The drafts are polite and accurate but sound nothing like him. Every one opens with "Thank you so much for reaching out", uses phrases such as "rest assured", and closes with "Is there anything else I can help you with?" His regular customers notice. One day he forgets to edit a draft before sending it, and a regular replies: "Daniel, is this a bot now?"
If you describe your voice with adjectives such as "friendly, down-to-earth", the assistant gives you its own generic version of friendly and down-to-earth. When you paste in 2 or 3 real pieces you wrote, it can see your actual sentence length, word choice, level of formality, use of contractions, and the way you open and sign off.
The samples should match the task. For customer replies, use customer replies. A LinkedIn post or a condolence message is still your writing, but it does not show how you answer a customer asking about delivery dates. Choose pieces you were happy with and that needed little editing, because those are closest to your natural voice.
Before you paste anything, remove names, order numbers and other confidential details. The model only needs your voice. It has no use for customer data.
Then give a specific instruction. "Match the sentence length, vocabulary and level of formality of these samples" works better than "write like me", because it names the features to copy and leaves the content to the new task.
Most people skip this step. Before you ask for any draft, ask the assistant to describe the style of your samples as a short list, for example: "describe the writing style of these three emails in five bullet points". What comes back is a style description: a short list, written by the assistant and corrected by you, of the style it sees in your samples. Asking for it first does two things. It shows you what the model has picked up, and it lets you correct it before any drafting starts.
Daniel gives the assistant 3 samples and asks for a description. It comes back with "casual, with frequent humour". He does not see himself as funny. Looking again, he realises one sample was a joking reply to a customer he has known for years, and the model treated it as typical. A single untypical sample can skew the model's reading of your whole style. He replaces that sample with a more ordinary reply and corrects the description to: "friendly but brief, no jokes with customers I do not know".
Once the description is right, save it. A few lines are short enough to paste into any request. Later you can put it into custom instructions or a project so it applies automatically; module 6 covers how to do that.
Sometimes the model keeps sliding back into a habit even with good samples. In that case, add a counter-example: a short piece of text, clearly labelled as what not to do, used to steer the model away from a tone or habit. The label is what makes it work. Without one, the model may treat the counter-example as another sample to copy. With the label, it sees a clear contrast between the samples to match and the text to move away from. Keep it short, since one paragraph is enough, and make sure it really shows what you dislike.
Daniel adds a counter-example to his prompt. It starts with a label saying this is the tone he does not want and the assistant should avoid it, followed by one of the over-polished drafts the assistant had written for him earlier.
Often the most reliable fix is to name the specific habit and tell the model to stop. Vague instructions such as "less formal" or "more natural" leave the model guessing. A named habit gives it something to check its draft against. Common ones to name are opening with a compliment, repeating the question back before answering it, ending every email with a question, stacking three adjectives, and adding a summary paragraph that repeats everything already said.
To find your own habits, read 3 or 4 drafts the assistant wrote for you and note what made you wince. Daniel ends up with a list of 4 instructions (paraphrased here):
Do not open by thanking the customer for reaching out. Do not end with a question unless I need an answer. Never use "rest assured", "kindly" or "we apologise for any inconvenience". Use the customer's first name once, at the start, and nowhere else.
Daniel's final request contains his corrected style description, 2 sample replies, 1 labelled counter-example and his list of habits to drop. He still checks every reply for accuracy before sending it, but he rarely needs to rewrite one for tone. Yours can follow the same pattern. Pick 2 or 3 matching samples you were happy with, strip out confidential details, and ask the assistant to match their sentence length, vocabulary and formality. Ask for a style description first, swap out any untypical sample, correct anything that is wrong, and save the result. Add a labelled counter-example if a habit keeps returning, and list the habits you winced at as specific "do not" instructions. Put all of it into one request, and keep checking the drafts for accuracy.
You start with an honest look at your own writing. Go to your sent folder and pick three emails you wrote quickly to people you deal with regularly, the kind you would be comfortable having an assistant learn from.
Paste three of your own emails and ask the assistant to describe your style in five bullet points. Correct anything it got wrong, then save the final description.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).