How to brief an AI so your LinkedIn draft sounds like you
An AI draft sounds like you when the AI was told six things; most prompts supply one. This guide gives you the full six-part brief, with a worked example. The six: who is reading, what the point is, what actually happened, how you write, what shape you want, and what to do when it is unsure. Half of the brief is written once and saved, half is new for each post, and it works in ChatGPT, in Claude and in any tool that takes instructions in your own words.
Reader: one person in one situation
Name the reader the way you would describe them to a colleague: “a practice manager at a dental clinic who does the marketing on Friday afternoons”, not “healthcare professionals”. Add what that person already believes about your subject; a post that only tells readers what they already think gets skipped. This part rarely changes, so it belongs in the half of the brief you write once.
Point: a sentence you would defend
Write the post's claim as one sentence before any drafting happens: “Most clinics lose new patients at the booking step, not in their advertising.” If you cannot, the AI cannot either — it will write around the subject and call that a post. A good point is one a reasonable reader could disagree with. It is new every time, and it is the part most often left out, which is why so many drafts explain a topic instead of saying something about it.
Material: what happened, in rough
Give the facts the post rests on, unpolished: what you saw, the figure and what it measures, what someone said, what you tried first. Bullet points will do; so will a paragraph pasted from an e-mail with the names taken out. Include the awkward detail — the mistake, the cost — since that is usually what readers remember. Whatever is missing, the AI will either leave vague or make up, and vague is the better of those two outcomes.
Voice: samples first, rules second
Paste two or three posts you wrote yourself, labelled as examples of how you write. Then add a few rules a draft can be checked against: your spelling, what you call your customers, whether you use emoji, the phrases you never want to see. Examples carry what rules cannot — sentence length, how you begin, how blunt you are — and Anthropic's prompting guidance singles them out as one of the most reliable ways to steer tone. Vague adjectives such as “authentic” add nothing; a contrast like “blunt, never sarcastic” is the most an adjective can do.
Shape: length, first line, ending
Say what the finished post looks like: roughly how long, a first line that stands alone and stays under about 80 characters, short paragraphs with blank lines between, and how it ends — on the point, on a number or on one specific question, not on “Thoughts?”. LinkedIn folds a post after about three lines, so an instruction about the opening does more than one about total length. The reasoning is at media-pilot.io/guides/linkedin-post-length.
Process: what to do when unsure
The last part governs behaviour rather than content, and it prevents most bad drafts. Three lines are enough. “Before drafting, ask me up to five questions about anything the material does not cover.” “Do not add facts, figures or quotations that are not in my material; where one is needed, write [CHECK].” “Give me two versions that open differently.” The second matters most: OpenAI and Anthropic both say plainly that their models can produce convincing statements that are not true, and a visible gap is easier to fix than a plausible invention.
Put it together: a worked example
Here is a whole brief for one post. “Reader: practice managers at dental clinics, who assume more advertising means more patients. Point: most clinics lose new patients at the booking step, not in the advertising. Material: we went through 40 enquiries for one clinic last month; 23 arrived outside opening hours; the booking form had eleven fields; they cut it to four and added a call-back option; bookings from the same advertising went from 9 to 15. Voice: as in the three posts above; British spelling; ‘patients’, never ‘leads’; no emoji. Shape: 900 to 1,200 characters, first line under 80 and on its own, end on the number. Process: questions first; no facts beyond these; two versions.” The figures are made up for the illustration. Yours come from your own work.
Save the half that does not change
Reader, voice and shape are the same next week; point and material are not. Keep the constant half where your assistant finds it without being told — custom instructions or a project in ChatGPT, a project's knowledge and instructions in Claude — and bring only point and material to each new chat. A LinkedIn tool keeps the constant half as settings: in MediaPilot, the company profile and the post rules — preferred wording, prohibited terms, tone and audience — are set once per workspace, and every generated post starts from them.