← Blog
Posts

Write down how you sound before an AI drafts for you

A one-page voice guide, written in an afternoon from your own posts and calls, gives ChatGPT or Claude something to copy and gives you something to check its drafts against.

Ghaith Ayadi · 20 Nov 2026 · 6 min read

A vague instruction gets the average back

"Make it sound like me" is the instruction that gives a model the least to work with. ChatGPT has never sat in on a sales call with you, never read the post that got the replies, and has no idea which words make you wince. So it falls back on what it was built for. OpenAI trained ChatGPT to predict the likeliest next word, and the likeliest word is the safe one a thousand other blogs already used. What ChatGPT hands back is polished. It could run under any logo, and you spend the afternoon rewriting it.

There is a measurement of how hard that pull is. A 2025 SSRN study, as ContentGrip reported, looked at businesses that lost access to ChatGPT during a temporary ban and found their content grew 15% more lexically diverse and 12% more syntactically varied. Those businesses lost the tool and changed nothing else. We think the fix is smaller than giving up the tool. ChatGPT writes like everyone when you have told it nothing about yourself, and the cure is one page you can check its drafts against.

Ten published sentences hold most of the voice

The voice is sitting on the blog already. Go through the posts and copy out ten sentences you would say out loud to a customer without changing a word. Paste them into one document. Then write down what they share: how many words they ran to, which words came up twice, where the jokes landed and who they were aimed at.

Our own site shows what turns up. "Docs rot, as our site puts it." ran to two words. "No reason, no pitch." ran to four, and "Revolutionary, apparently." took a swing at the industry in two. When we put those lines side by side, we learned more from them than from any adjective. They said we write in fragments, we drop the verb when it slows things down, and our jokes go at the trade. In the voice guides we have read, the same four adjectives turned up near the top, friendly, confident, expert and human, and any company with a website could have claimed all four.

Write the bans before the preferences

You will have a hard time checking a mood. "Warm but authoritative" needs someone to judge it, and your judgement shifts with how your week went. A banned word takes ten seconds and the Find key. Our house rules open with four bans. No em dashes, no sentences that deny one thing and then correct it instead of comparing the two, no "in today's fast-paced world", and no closing paragraph that repeats the post. We can hold any writer to those, and any model too.

Start the page the same way. Write down the phrases that have never once come out of your mouth and still keep turning up in AI drafts, and put them at the top.

Calls catch the voice before anyone edits it

Published posts went through an editor, even a tired one at midnight. Your calls didn't. On a sales call the same point gets made many times a week to one person at a time, and after enough repetitions the phrasing settled into something that worked on a stranger. Open one transcript from a call that went well and take two phrases you would happily see in a post. A comparison that landed, a blunt answer to a hard question, the way the problem got described before the prospect could.

Those two phrases go on the page as the unedited sound of the company. Our rule for calls holds here too. We keep names, companies, budgets and vendors from the call out of every post, and take only the wording.

Every rule gets a before and after

When ChatGPT gets a description of a voice, it interprets it, and its interpretation drifts back toward the average. With an example, you leave less room for that drift. So each rule on the page carries one sentence written the wrong way and the same sentence fixed, and both should come from your own drafts. For our ban on the contrast trick, we would write the before as a sentence that denies one thing and then names another instead of comparing them. The after is closer to how our site puts it, that writing is the easy part and knowing what is true takes longer.

Marketers seem to agree that a person has to stay on the voice. In PhotoShelter's 2026 survey of nearly 400 marketing, communications and creative professionals, 96% said human oversight remained essential to originality and brand voice. We'd add that the oversight goes faster when the human can point at a line on a page and say "that one".

Keep the guide to a single page

Docs rot. A brand book is right the day it gets written, and then it sits in a shared drive while the company changes around it. If you write forty pages, they rot faster, since you won't open a forty-page file to settle whether a joke went too far. One page fits in a prompt, fits on a screen next to the draft, and gets read by a freelancer before the first paragraph. If the guide needs a table of contents, it has outgrown its job. Cut the rules nobody ever broke and keep the ones that keep coming up.

Add a rule the second time you fix something

The first time a fix happens, it is a fix. The second time the same thing gets fixed, it is a rule, and it goes on the page that day with the before and after from the draft that prompted it. Waiting for the end of the month means the example is gone and the rule comes back as an adjective.

We ran into this on our own blog. On a recent draft, our reviewer cut a sentence claiming that nearly every company that ever paid for a voice guide got the same adjectives. In its place our reviewer wrote that every voice guide we had read used them, a claim we can stand behind. If that edit comes up a second time, it becomes a line on our page about claiming only what we have seen.

Rewrite the page every quarter

A rulebook can suck the life out of anything. Followed long enough, the rules that once kept posts sharp start making them sound alike, and sameness is what the guide was written to escape. The same PhotoShelter survey found that 45% of marketers still struggled to stand out from competitors, a share unchanged since 2023, and that 80% worried generative AI was filling the internet with generic, repetitive content. A guide that hardens into a template adds to that pile.

Orwell saw this coming too. In his sixth rule he told writers to break any of the other five sooner than say anything outright barbarous. So once a quarter, read the page against the last ten posts. Merge the rules that overlap, delete the ones that never came up, and loosen whichever one made three posts in a row open the same way.

How Propaganda's voice guide fits in

Propaganda's voice guide is in beta. It writes a one-page guide from the posts a company has already published and suggests changes drawn from the edits that keep repeating, which is the method above with the counting done by a machine. The quarterly rewrite is on our roadmap and is not built yet, so for now that part stays with a person and an afternoon. Write the page either way. The model will draft just as fast tomorrow, and this time it will have something to copy.