The hard part of writing in someone else's voice is getting their real opinion onto the page. Here's a 15-minute capture, a draft prompt and an approval step that keeps the founder in charge of what's said.
Don’t write a founder’s posts from their profile. Record five questions, paste the transcript into AI with a prompt that forbids invented detail, and have it list every claim so the founder can approve the facts in two minutes. The founder keeps the opinions, the numbers and the comments. You keep the calendar, the editing and the record of what was said.
Say you’re the marketing manager at a 30-person software company. Your founder, Ravi, wants to be more visible on LinkedIn. Over a month you’ve drafted three posts for him from his bio, the company blog and the brand deck. He sent two back with “not me” and one with “fine, I guess,” which he never posted.
Then on Thursday at ten to seven he sends a four-minute voice memo from the car. It’s rambling. It’s about a customer who left because onboarding asked them to fill in a spreadsheet nobody wanted to fill in, and he says, with some feeling, that he’d been blaming the product when the problem was the first week. You play it twice and realise it’s worth more than all three drafts put together, because it contains an opinion, a detail and a sentence only he would say.
That’s the whole idea of this guide. When you write in someone else’s voice, the drafting isn’t the bottleneck. The source material is, and AI is very good at drafting and has no access to what the founder believes. So the order of work matters: capture the founder’s own words first, let AI shape them into a post, and have the founder approve what the post claims.
The drafting half of that is well supported. In a 2023 experiment with 444 college-educated professionals writing things like press releases and delicate emails, using ChatGPT cut the time taken by about 37%, and graders rated the output higher.[1] Notice what those participants were handed, though: a task with the brief already decided. Ghostwriting hands you the opposite, a blank brief and a person who’s busy, which is why the time you save on drafting only pays off if you spend some of it on the interview.
The capture is a recorded conversation of about 15 minutes, once a week, on the founder’s own phone or in a call they start themselves. A standing slot on their calendar works better than a request each week, because a request can be postponed and a recurring slot usually turns into a habit. You ask five questions, in any order, and you let them ramble.
| Question you ask | What it gets you | What Ravi said (shortened) |
|---|---|---|
| 1. What did you change your mind about recently? | A position with something at stake | “I used to treat onboarding as a product problem. After last month I think it’s mostly about who has time in week one.” |
| 2. What did a customer, a hire or a competitor do this week that surprised you? | A scene, which is what makes a post feel lived in | “A customer left because we asked them to fill in a spreadsheet nobody wanted to fill in.” |
| 3. What do most people in your industry get wrong? | A contrarian angle, in their own words | “Everyone ships more features. Nobody checks whether the customer has a free afternoon to set them up.” |
| 4. What would you tell someone starting in your seat? | Advice that sounds like them, not a listicle | “Sit in on three onboarding calls before you touch the roadmap.” |
| 5. Which numbers or facts can I safely quote, and what’s off limits? | The only figures AI may use, and the topics to stay away from | “Onboarding takes customers about three weeks. Don’t name the customer. Don’t mention the funding.” |
A worked example for a software company founder. The answers are illustrative, not from a real interview. Question 5 is the one people skip.
Question 5 is the one teams leave out, and it’s the one that saves you later. Without it, the AI will supply numbers that sound right, and you’ll find out in the approval step at best. With it, you have a short list of figures the founder has said out loud, and everything else in a draft can be treated as unverified.
No post starts from a blank page or a profile. It starts from something the founder actually said.
A few practical points before the first recording:
Now the drafting. The mistake is pasting the transcript into a chatbot with “turn this into a LinkedIn post.” What comes back is fluent and smooth, and it has quietly replaced the founder’s odd, specific phrasing with the phrasing every other post uses. Worse, it will add claims: a statistic, a lesson, a confident “here’s what I learned” that nobody said.
Choose your tool with one rule. Can you paste in a transcript and two or three of the founder’s real posts or emails, and get back a draft that marks which lines came from the transcript? If yes, it will do the job, whichever assistant you already have approved at work. A template library, or a “write in my voice” button that doesn’t see the transcript, is solving a different problem. The prompt below is what does the work, filled in for Ravi’s week.
| You’re helping me draft a LinkedIn post for our founder, Ravi, who will read and approve it before it goes out under his name. Use only what’s in the transcript below. Keep his own words and phrasing wherever possible, including any odd or blunt ones. Format: 150 to 220 words, first person, short paragraphs, no hashtags, no emojis, no “Here’s what I learned”, no closing question. Open with the specific moment from question 2, not a general statement. Rules: Do not add any fact, number, name or example that isn’t in the transcript. Do not name the customer. Do not mention funding. If a point needs a number and the transcript has none, write [NEEDS NUMBER] instead of inventing one. After the post, give me a list titled “Claims to approve” with every factual statement in the post, each followed by the exact line from the transcript that supports it, or “no source” if there isn’t one. Two samples of how Ravi writes: [paste two short real posts or emails] Transcript: [paste transcript] |
Swap in your founder’s name and format limits. The “Claims to approve” list is what makes the approval step fast.
Read what comes back against the transcript, not against your memory of the founder. Then edit, and expect to make cuts more often than additions. Here’s a typical pass on a first draft for Ravi’s post.
| The first draft said | Source in the transcript | You change it to |
|---|---|---|
| “In today’s fast-paced SaaS world, onboarding is everything.” | None | Cut. Open with the customer who left. |
| “We lost a customer because we asked them to fill in a spreadsheet nobody wanted to fill in.” | Question 2, word for word | Keep exactly as he said it. This is the line that sounds like him. |
| “After we simplified onboarding, churn dropped 30%.” | None. No figure was given and nothing was changed yet | Cut the sentence. A claim about a result that hasn’t happened is a damaging kind to get wrong. |
| “Sit in on three onboarding calls before you touch the roadmap.” | Question 4 | Keep, and make it the closing line in place of “Agree?” |
| “It taught me an important lesson about customer empathy.” | None | Cut. He said “I’d been blaming the product,” so use that instead. |
Illustrative first-draft lines of the kind a model tends to produce unprompted, checked against the example transcript above.
Two habits make this pass faster. Keep a running file of the founder’s phrases, such as the things they say about customers, the words they overuse and the ones they won’t touch, and paste it in next to the samples. For a wider process for building that file, our guide to training AI on your brand voice applies, with one change: a founder’s voice includes opinions the company voice doesn’t.
Many founders, when sent a draft, read it for tone. “Does this sound like me?” is a natural question, and it’s a poor check, for a reason that surprised me. In a 2023 set of experiments, people asked to tell AI-written self-descriptions from human-written ones scored 50 to 52%, which is about what guessing gets you. They leaned on cues that didn’t separate the two, such as grammar slips and long words.[2]
What that means for your approval step is that a founder’s ear isn’t a reliable detector, and neither is yours. A post can sound exactly like Ravi and still say something he didn’t say or wouldn’t stand behind. So change what you ask him. Send the post and the “Claims to approve” list, and ask one thing: is each of these true, and would you say it to a customer’s face? That takes two minutes, and it protects him better than a read for tone does.
Ask the founder to approve what the post says, not how it sounds.
Polish is your job and the claims are theirs. Writing the split down stops each side assuming the other is checking.
| What AI does | What the human still owns | How it gets checked |
|---|---|---|
| Turns the transcript into a first draft in the founder’s phrasing | The opinion itself, and whether it’s still what the founder thinks | The founder answers “true, and would you say it to a customer?” on each claim, before publishing |
| Lists every factual claim with its source line | Any number, customer detail or result; only the founder can confirm these | Marketing manager checks each “no source” line and removes it or gets it confirmed |
| Suggests two or three alternative openings | Choosing the opening, and keeping the founder’s real phrases | Marketing manager compares the final post with the transcript once, for wording that wasn’t said |
| Drafts a holding reply for comments the founder hasn’t answered | Every reply and every comment left under the founder’s name | Founder posts replies from their own account, weekly, in five minutes |
| Nothing, for posts about a named customer, employee, funding or legal matter | The decision to post at all | Founder and the person named, or legal, approve in writing |
A starting split for a two-person setup. Keep the last row even if the company is tiny.
One more limit. If you’re asking whether AI should help with a sensitive post at all, our fit test for when not to use AI at work is a quick way to decide. Posts reacting to a layoff, a lawsuit or a customer’s complaint tend to fail that test, and a founder can write two sentences on those faster than you can brief an assistant.
Here’s the rhythm put together. The founder’s part is deliberately small: a recording, a two-minute approval and their own comments. Everything else sits with you.
The capture, draft, approve loop. The founder’s weekly time is roughly 20 minutes; the rest is yours.
Friday’s note is the step that turns a few good weeks into a system. Every edit the founder makes is a clue to how they sound or what they won’t say, and pasting those notes into next week’s prompt is what makes “not me” become rarer. Keep four things together in one folder:
Now picture the week going wrong. Ravi records his memo, and in the draft the assistant writes that onboarding “took customers three weeks on average, down from five.” He said it felt like three weeks. Nothing in the transcript mentions five. The claims list shows “no source” next to it, you ask him on Wednesday, and he says, “No, that’s a guess, cut it.” Without that list, a number that started life as a feeling in a car would have been published under his name, and a prospect who knew better would have noticed first.
Before anything goes to the founder, score it yourself.
| # | Question | If the answer is no |
|---|---|---|
| 1 | Does the opening come from something the founder said this week? | Go back to the transcript and find the moment. Don’t open with a general statement. |
| 2 | Is every number in the post one the founder gave you? | Cut it or mark it [NEEDS NUMBER]. |
| 3 | Is there at least one phrase in it that only this founder would use? | Put one back from the transcript. |
| 4 | Is anyone named or identifiable who hasn’t agreed to be? | Remove them, or get their agreement first. |
| 5 | Does the post end where the founder’s own point ends, without a question added to farm comments? | Trim to their last real sentence. |
Scoring: five yeses, send. Four, fix the one and send. Three or fewer, go back to the transcript before the founder sees it.
The founder posts, replies and answers messages from their own account. You prepare, they publish.
I’d hold that line for two reasons. Comments and direct messages are where people test whether there’s a person behind the post, and an assistant-written reply under a founder’s name is a fast way to fail that test. And logging in as someone else creates a security and account-terms question that a scheduling note or a shared draft avoids entirely. LinkedIn’s current terms on account access were behind a block for automated readers when I tried to check them, so read them yourself before anyone logs in as someone else.
“The founder” is a convenient label, and the workflow flexes a good deal depending on who’s writing for whom. The capture stays the same. The risk and the approval step shift.
| Who you are | Whose voice | Best way to capture | Biggest risk | Approval step |
|---|---|---|---|---|
| Solo marketer at a small company | Founder, who has little time | A recurring 10-minute voice note | Drafting from the website because the founder is quiet that week | Claims list in a message thread, answered in two minutes |
| In-house comms manager | CEO, with an executive assistant | Short call booked by the assistant, transcript shared | Policy, regulatory or investor-sensitive wording slipping in | Comms lead plus legal review on anything touching results, hiring or funding |
| Freelance or agency writer | Several clients at once | Client-recorded note, one per post | Drafts from different clients converging on the same phrases | Separate voice file per client, client sign-off on the claims list |
| Content lead | A subject expert, such as a head of engineering | Thirty-minute conversation, then split into several posts | Technical errors that read confidently | The expert checks the facts; you check the clarity |
| The founder, writing their own | Themselves | Talk the answers into a phone, transcribe, then draft | Skipping the approval because it’s “their own” post | Reading the claims list the next morning, not the same evening |
Setups differ by role. Adapt the approvers to your size and add anyone your industry requires.
It’s fair to ask whether the effort is worth it. In Edelman and LinkedIn’s 2025 survey of 1,934 business executives, 64% of hidden decision-makers said they trust thought leadership more than marketing materials and product sheets when judging a supplier’s capabilities.[3] That’s about thought leadership broadly, not LinkedIn posts in particular, so take it as a reason to care about substance rather than a forecast for your posts. The study doesn’t cover who wrote them, which is an argument for making sure the substance is the founder’s.
On the ethics question that comes up in every workshop: ghostwriting for executives is an old practice, and adding AI doesn’t change whose name it is. What changes is the responsibility. The person whose name is on the post has to stand behind every claim in it, which is why the approval step is built around claims. Some companies and industries have their own policies on AI-assisted external writing, so check yours. And if your founder would be uncomfortable being asked “did you write this?”, have that conversation before the first post.
Posts going out on schedule tell you the process is alive, and they tell you nothing about whether the posts are any good. The measurement lens we use at Future Factors has three steps: use, persistence and impact. Here’s how they look for a founder’s LinkedIn.
| Step | What to pull | Who pulls it, and when | What it tells you |
|---|---|---|---|
| Use | Posts published against posts planned, and minutes of founder time per post | Marketing manager, monthly | Whether the loop is running and whether it fits the founder’s week |
| Persistence | In month three, are captures still happening without a chase, and is the founder still answering comments personally? | Marketing manager, month three | Whether the habit survived once the novelty wore off |
| Impact | How often the founder changes “what it says” (not wording) in a draft; conversations sales or customers start from a post; how often the sales team reuses a post in follow-ups | Marketing manager with the sales lead, quarterly | Whether posts carry real opinions, and whether anyone outside marketing finds them useful. These are proxies, so read direction, not proof |
A measurement sheet for a founder’s posting. Reach and impressions are quick to pull but say little about whether the posts are working.
It also helps to keep the parts of change apart. Showing a marketing manager the prompt and the claims list is training, and it builds capability. The weekly loop gives that capability a recurring place to land, which is workflow design. The Friday log and the month-three review are reinforcement, the part that turns it into behavior. Adoption is what you see if all three happened: the founder still recording in month four because it’s become part of the week. Without reinforcement, teams tend to drift back to drafting from the website.
Teams can start this internally with the tables above, and plenty will. Hands-on training shortens the learning curve in specific ways: it lets the marketing manager and the founder practice on a real transcript, it gets a whole comms team working to the same approval standard rather than one person who’s good at it, and it covers the judgment calls, like what to do with a sensitive post, that a guide can only point to. If you also run LinkedIn for the company page or for several leaders, our guides to employee advocacy on LinkedIn and to writing your own LinkedIn posts with AI cover the neighbouring jobs.
Here’s a first step for this week. Book one 15-minute slot with the founder, ask questions 1 and 5, record it, and draft a single post with the prompt above. Send it with its claims list, and keep the founder’s edits for next Monday.
Executives have used writers for a long time, so the practice itself isn’t new. What matters is that the views are the founder’s, that they’ve read and approved every claim, and that nothing is presented as something they did or said when they didn’t. If your company, industry or the platform has rules on AI-assisted writing, follow those, and decide with the founder in advance how they’d answer if asked.
Old posts help with phrasing, and two or three samples in the prompt are worth having. They can’t supply this week’s opinion, and they tend to lock the output into whatever the founder wrote before, including habits they’d like to drop. Use the samples for voice and the interview for content.
We’d start with one or two and raise it only once the capture, draft and approval are running without chasing. A rhythm the founder keeps up for six months is worth more than a burst they can’t sustain. The right number depends on the founder’s time and how many real opinions they have, so treat two a week as a starting point, not a benchmark.
Shrink the ask before you skip it: a five-minute voice note answering two questions, or a recording of a call the founder was going to have anyway, with the other person’s agreement. If you truly get nothing, that tells you something about whether the program has the founder’s commitment, and it’s better to pause than to invent opinions for them.
No. Those replies are where people find out whether there’s a person behind the post. Let AI draft a holding line or summarise a long thread if that helps, but the founder replies from their own account. If they can’t spare five minutes for comments, publish less often.
The 37% time figure comes from the author-hosted paper by Noy and Zhang at MIT (444 professionals; the version published in Science, 2023). The detection accuracy figure is from Jakesch, Hancock and Naaman in PNAS (2023). The 64% figure is from page 13 of the 2025 Edelman-LinkedIn B2B Thought Leadership Impact Report (1,934 executives, surveyed March 17 to April 3, 2025). All three were read at the publishers’ or authors’ own pages in September 2026. LinkedIn’s terms on account access could not be read by an automated reader, so this guide makes no claim about them. Posting-frequency and reach statistics from ghostwriting vendors were left out because none traced to a primary source. The founder, company and numbers in the examples are illustrative, not real.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.