AI ghostwriting works for executives when the AI learns from evidence of how you actually write. Real voice samples, a written voice skill with hard rules, a ban list and a few before/after pairs replace self-description. First drafts then come close to your voice, and a feedback loop that turns each correction into a rule closes the gap.
Why does AI-written content sound generic?
Because the AI writes in its own default voice unless it gets precise instructions. That default voice is fluent, balanced and interchangeable. Readers notice it, and platforms increasingly do too.
Most executives who try AI ghostwriting describe how they think they sound: "direct, warm, strategic". The AI cannot act on that. It needs evidence of how you actually write and rules specific enough to change its output.
I use Claude for my own writing, and this guide describes how I set it up. The principles carry over to other tools.
Where do executives get voice samples for AI?
From texts and recordings that already exist. Your voice is in years of emails, posts, transcripts and messages. Ranked by how much they reveal:
- Published texts you wrote yourself: posts, articles, newsletters. This is your deliberate written voice. Take the pieces where you read them back and think "that is exactly me". Leave out anything ghostwritten or heavily edited by others.
- Sent emails: your direct, unpolished voice and its range, from peers to clients to your team.
- Meeting transcripts: how you open a point, how you disagree, the phrases you reach for without time to polish.
- Podcast and keynote transcripts: your performed spoken voice, more structured than a meeting, more spontaneous than writing.
- Messages and comments: the least filtered layer. Many of your most characteristic phrases live here.
- Interviews and speeches: how you phrase things under questioning, and your voice at its most deliberate.
Fifteen to twenty-five samples across at least three source types are a solid base. Quality matters more than volume: is this me at my best, or me on autopilot?
Two practical rules. First, mark each sample as written or spoken. The two voices differ, and the AI can only separate them when it knows which is which. Spoken material is rich in vocabulary and rhythm, and the written samples set the structure. Second, extract the style and leave the substance behind. Emails and transcripts contain names, deals and numbers. Anything that ends up in the instructions gets anonymised first.
Then let the AI analyse the samples and compare the result with how you would describe your own voice. Where the two disagree, you decide which version it should write in: the one the evidence shows or the one you want to grow into.
What is a voice skill?
A voice skill is a written instruction that makes the AI write like you in every draft. It differs from a style guide in one way that matters. A style guide is written for people. A voice skill has to work for a machine.
So each line has to pass one check: would the AI write anything differently because of this line? "I am a passionate entrepreneur who thinks in systems" changes nothing. "Never open with a rhetorical question" changes every opening.
My own voice skill has this shape:
- When it applies: posts, articles and newsletters yes, legal documents and technical specs no.
- The voice in one paragraph: specific enough that an outsider would recognise the texts from it.
- Hard rules: what never happens, written as prohibitions.
- A ban list: words and phrases that never appear.
- Before/after pairs: the same statement once generic and once in the executive's voice.
One rule at the very top breaks every tie. Mine is clarity over completeness. If elegance and clarity pull in different directions, clarity wins. Without a declared winner, the AI resolves conflicts at random.
One hard rule belongs in every executive's skill: never invent or alter numbers, facts, names or quotes. Every figure comes from your input, and a missing figure gets a placeholder. Style is open to taste. Facts are not.
Which voice decisions should an executive make?
Eight decisions define most of a written voice. The voice samples decide them where they can:
- Directness: thesis first, a short scene first, or a question first.
- How a point travels: argument, story or evidence in front.
- Tone: factual, warm, sharp or referential, with one dominant register.
- Humour: none, an occasional dry beat, or woven through.
- Personal presence: behind the subject, own experience as material, or own mistakes in public.
- Vocabulary: the ten to fifteen words people in your field actually use, as shown by your samples.
- Ban list: motivational clichés, press-release buzzwords such as "holistic" or "game-changer", interchangeable openers such as "In today's world", and your personal irritants.
- Mechanics: sentence length, punctuation, paragraph length, how you address the reader, emojis and bold text.
One item deserves its own place on the ban list: the hollow antithesis, as in "It's not just X, it's Y", where the negation corrects nothing the reader believed. In my experience, banning it is the fastest way to stop sounding machine-written. The fix is to say what something is.
Why do before/after pairs matter most?
Because the AI learns more from a difference it can see than from an abstract rule. A pair shows the same short text twice: once correct and interchangeable, once in your voice.
A constructed example. Before: "Strong investor relationships are essential for a successful fundraising process and should be built early." After: "The round closed in six weeks. The relationships behind it started two years earlier, over coffee with people who had no reason to invest yet." The first could come from any report. The second has a person behind it.
Three strong pairs are enough. Each gets a one-line comment that names the rule behind the change, such as "thesis moved to sentence one". The best material comes from the passages in your samples where you sound most like yourself. You write the generic version next to them.
How good are the first AI drafts in your own voice?
In my experience, the first drafts land at roughly 70 to 80 percent. That is the right result at that stage. A feedback loop closes the rest.
Executives who use AI already correct drafts, with notes like "not my style" or "too salesy". Normally those corrections vanish when the chat closes, and the same mistake returns the next day. The loop catches it. When a correction names a new mistake, the AI asks whether it should become a rule, then adds it as a pair or a hard rule to the skill. Known mistakes it simply avoids.
A real case from my own work. The AI once fixed a passage by opening with a negation, "The model does not mean X", and planted the exact wrong idea it then tried to remove. I replied with one line: "Why do we say what it isn't?" Since then, my skill carries a rule: the fix is almost always the simplest positive statement.
One routing rule keeps the skill clean. How you sound goes into the skill. What works on a platform, such as hook mechanics, formats and posting times, belongs in a separate playbook.
Does AI ghostwriting replace the executive's judgment?
No. The engine produces drafts, and you decide what goes out.
What changes is the reach of your voice. Many texts a leader sends never appear on LinkedIn: the email to the leadership team before a hard decision, the board update, the all-hands, the keynote, the message to a candidate. Written under different time pressure, often by different hands, they can make one leader sound like three people. With a voice skill, every one of them can start in the same voice and carry the same positions. My guide on CEO communication covers why that consistency matters inside the company as much as outside.
“Without a declared winner, the AI resolves conflicts at random.”
Frequently asked questions
Can AI write LinkedIn posts in an executive's voice?
Yes, when it learns from real samples of how the executive writes and works from a voice skill with hard rules, a ban list and before/after pairs. A description like "direct and warm" alone is not enough.
How many writing samples does AI need to learn my voice?
Fifteen to twenty-five strong samples across at least three source types, such as published posts, sent emails and one spoken source, make a solid base.
Is AI ghostwriting risky for executives?
The main risks are invented facts and a generic voice. A hard rule against inventing numbers, names or quotes, and a final review by the executive before anything goes out, address both.
Do you help executives set up AI writing in their own voice?
Building a CEO's voice and the system that carries it is part of my executive branding work. You can read more about my executive branding advisory.
