An article that “sounds like a machine”, and a client who notices. Over one week, an agency copywriter teaches AI to write like her, without exposing the unpublished texts entrusted to her.
“That one reads like it was written by a machine.” The health insurer's communications manager said it without malice, in a meeting. Inès, a copywriter at a small content agency, said nothing: the article had indeed been drafted with a consumer AI tool. Smooth sentences, “truly” three times. Nothing wrong in it, and nothing of her.
Worse: to get that draft, she had pasted in the client's brief, including the figures for an offer that would not be announced until the following month.
Here is how Inès, a fictional copywriter, took back control the following week.
A copywriter sees an offer before its launch, figures before they are published, the opinion piece an executive will sign with their own name. The service contract often includes a confidentiality clause, and a company's unpublished information is protected as a trade secret in most countries, Switzerland included. Pasting a brief into a consumer tool means handing it to a third party the client never agreed to.
If the agency or the client forbids AI, that rule applies: raise it with them, don't work around it.
Inès picks three articles already published for the health insurer. She opens IA Confidential, stays with the confidential models offered by default, and writes the request she will reuse for every client. The first example to copy:
Here are three texts I wrote for the same client. Describe my style as concrete, checkable rules: average sentence length, punctuation, recurring vocabulary, how I open and close, how I address the reader, and above all what I avoid. No compliments and no vague adjectives like “fluent” or “engaging”: only rules that could be checked against a new text.
The answer comes to twelve rules. She strikes out two that are wrong and adds what the AI could not have guessed: never “unmissable”, no bullet points, always an example in the first paragraph.
She creates a “Health insurer” space, using the Writing type. In the space's Assistant tab, she pastes her corrected style sheet. These instructions apply to every conversation in the space, ahead of her general preferences. She adds one line of method: “When you are not sure of a fact, put it in square brackets instead of making it up.”
On her computer, with Chrome, Edge or Opera, Inès adds to the space the folder where she keeps every article delivered to the health insurer, ticking the “Confidential models” option: that folder is never sent to an anonymised model. When she asks “how did I present the cover for glasses and contact lenses last year?”, the assistant finds the relevant passages on its own.
In the space, she pastes her interview notes and asks:
From these notes, write a 600-word first draft that follows the space's instructions. Keep my figures exactly as they are, and put in square brackets anything you had to assume.
The draft is uneven, but it sounds like her. One problem: a launch date that appears nowhere in her notes. She removes it and rewrites the opening herself.
Inès no longer asks the AI to rewrite, but to review. This is the second example to copy:
Review this text against the space's instructions. Flag every sentence that strays from them, with the rule it breaks and the word causing the problem. Don't rewrite anything: I'll make the corrections myself.
Seven sentences come up, three of them written by her own hand on a tired evening. She corrects them.
The assistant lays the article out as a document, which she edits and then exports to Word. The layout is rebuilt from the text, so Inès moves it into the agency's template before sending. On Monday, the communications manager replies in one line: “Perfect, that's definitely you.”
A style sheet captures the surface: rhythm, words, habits. Not a point of view. Without a review, the AI slips back into its stock phrases and flattens any voice; it can also invent a figure or a date. The text is still signed by you or by your client: every fact has to be checked.
With the confidential models, the brief, the notes and the drafts are processed on servers in Switzerland and Europe, outside the reach of the CLOUD Act, and nothing is kept on our side. The history stays in your browser: we don't have it. A folder added with the “Confidential models” option is never sent to an anonymised model.
An anonymised model remains your choice, always behind the confidentiality filter, which detects sensitive data before anything is sent. For a text under embargo, stay with the confidential models. And since the history lives in the browser, anyone using the same session can read it: on a shared agency computer, work in a personal session and log out when you leave.
Before pasting a brief, read through the data you should never paste into an AI. Screenwriters look after their world the same way: see their guide.
To draw your own style sheet from two paragraphs you have already published: open the chat.
Try it on your own case No account needed · First month free
Open
