Writing with AI means giving an AI chat such as ChatGPT, Claude or Gemini the purpose of the text, the reader, your material and an example of your tone. In return you get a draft within seconds, which you check and approve. The AI writes quickly and cleanly, but whatever you do not tell it about your business and clients, it fills with something average or simply makes up.
I work with Claude every day. My fourteen AI employees run on Claude Team, some also on open models or through the API. Two of them mainly draft texts: Sonja drafts my business emails, Conny my Instagram captions. Nothing gets sent or posted that I have not read myself. All the basics about AI are collected on my AI Basics page.
What the AI needs from you to write
An AI chat runs on a language model. That is a program that has learned from huge amounts of text which words are likely to follow each other. This is why its texts read smoothly. Whether they fit you depends almost entirely on what you give it.
Anthropic, the company behind Claude, offers a good picture for this: think of the AI as a brilliant but new employee who does not know your norms and workflows. The more precisely you explain what you want, the better the result. Five pieces of information are enough for most texts:
What you provide
Why
Example
Purpose
Otherwise the AI does not know what the text should achieve
Clients should book their appointment before the holidays
Reader
Tone and wording depend on it
Regular clients you are on first-name terms with
Material
The AI cannot know facts, only guess them
Closed from 22 Dec to 6 Jan, emergency line on the mobile number
Tone
Without a model it writes neutrally and politely
two texts you wrote yourself
Format
Otherwise it usually gets too long
five sentences at most, no bullet list
The fourth row matters most. Anthropic writes that examples are one of the most reliable ways to steer the format, tone and structure of a response. A description such as "casual but professional", on the other hand, fits almost any text.
Step by step: from bullet points to a finished text
An example, made up, as it could happen in any small business: a painting company wants to tell its regular clients about the holiday closure.
Write down the message yourself in bullet points. What is new, what should the client do, by when? These three lines come from you, not from the AI.
Give the brief with the five pieces of information. Something like: "Write a short email to our regular clients. Purpose: they should book appointments before the holidays now. Facts: closed 22 Dec to 6 Jan, emergency line 0170 … Tone: like the two emails below. Five sentences at most. Do not invent anything that is not stated here."
Read the draft and make targeted changes. Instead of "make it better", say exactly what should change: "Put the emergency line in the first sentence", "no exclamation marks".
Check the facts against your bullet points. Dates, phone number, prices. This is where the expensive mistakes happen.
Approve it yourself. Only then does the text go out.
The last sentence of the brief in step 2 matters a lot to me. Conny has a fixed rule: she does not invent personal facts about me. If something is missing, the draft shows a visible gap instead of a plausible invention. I fill it in thirty seconds myself. Why an AI otherwise tends to make things up is explained in AI Hallucinations: Why AI Makes Things Up.
You do not have to rewrite recurring briefs every time. Ready-made briefs for quotes, client emails and texts in your own voice are in my prompt templates. A prompt, by the way, is simply the instruction you write to the AI.
Which texts suit the AI and which do not
Whether an AI draft works depends less on the type of text than on how much decision-making is in it. Routine works well; sensitive messages are better written by you.
Text
AI draft?
Where I cover it in detail
Business emails, appointment confirmations, standard answers
The last row has a simple reason: there, the text itself is the decision. If you have it drafted, you are not handing over the wording but the weighing up.
What to watch out for
The message comes from you. The AI improves language, structure and length. Only you know what you want to say.
Client data does not belong in the chat unchecked. Before I write real client data into a prompt, I replace names, amounts and company names with placeholders like <Client A> or <Amount>. Only once the template is set do I fill in the real details, and even then only as much as the task actually needs. More on this in the section on anonymising in Claude Prompts: Templates.
Strip typical AI traces. An opening line like "I hope this email finds you well", three adjectives in a row, enthusiasm without a reason. My full list is in Can AI Write Emails in My Voice?. Whether AI text can be reliably detected at all is covered in Can You Detect AI Text? What Detectors Do.
Avoid long chats. Ideally, open a new chat for each new text. Otherwise the AI mixes tone and facts from earlier tasks into it. Why that happens is explained in Context and Context Windows in AI Explained.
There is no fixed ranking, because the result depends heavily on the brief. ChatGPT, Claude and Gemini all write fluent text. My advice: give two tools the same brief with the same examples and pick the one whose draft you have to change less.
Can the AI write in my own tone?
Yes, if you show it how you write. Sonja writes my emails following a style guide built from 50 of my real sent emails: greeting, sentence length, sign-off and a list of things I never write. A description like "friendly and direct" is not enough on its own.
Do I need good prompts for this?
You need a clear brief, not a magic formula. The five pieces of information above are enough for most texts. If you keep needing the same kind of text, a fixed template that you only fill with your material is worth it.
Will readers notice that the text comes from an AI?
Sometimes, from typical phrases and missing details. With your material and your examples, the text becomes more concrete and sounds more like you. More important than whether anyone notices is a different question: is every detail correct, and would you sign the text as it is?
Can I send AI texts straight away?
I don't. A draft is a proposal, not a finished job. With me, approval always stays with me, even when the draft comes from one of my own AI roles.
Where to go from here
Take a text you have to write this week anyway and give the AI the five pieces of information from the table above, plus two of your own texts as a model. Then compare the draft with what you usually get. If you are just starting out with AI, read AI for Beginners: What to Know First first. The next step goes the other way: Summarize Texts and PDFs with AI. How others write and check their texts with AI is what we discuss in my community Claude Practitioners.
Kevin Welter
Developer, IT architect, author of technical books (Kubernetes, cloud infrastructures) and speaker. Runs his business with an AI workforce of fourteen AI employees and shows solo business owners in his community how to hire their first AI employee.