In October, a machine drafted an invitation for me, and in the middle of it sat a sentence that stopped me. It described the reader’s limit with a word-for-word Slovenian rendering of the English “ceiling”. Grammatically flawless. Foreign to any Slovenian ear. Under the draft I wrote: “Totally unnatural. We don’t talk like that.”
Then I went looking for the culprit and found it in my own instructions. A rule I had written myself asked the writer for exactly that word. The machine hadn’t made a mistake. It had obeyed.
So I didn’t fix the sentence. I fixed the rule.
Almost everyone noticed the speed. One in eight saw better results.
There is a name for this: “agentic writing”. Writing the rules that agents write by, and keeping them up to date.
This summer, the Content Marketing Institute surveyed 861 B2B marketers and found that 88% use AI to create content. Asked what it brings them, 78% chose speed. Only 12% chose better results.
Rory Sutherland would say we are building another Eurostar. Britain spent six billion pounds on new track to cut about 40 minutes off the journey to Paris. Sutherland suggested spending a fraction of it on hiring the world’s top models to pour free Château Pétrus for the passengers instead. People would ask for the train to go slower.
Readers don’t care how fast a text was produced. What they notice is whether anyone made the effort. When nobody did, the result already has a name: in 2025, Merriam-Webster chose “slop” as its word of the year.
Why only one in eight? My explanation: most people fix the machine’s sentences. A good writer ends up spending all the time saved on corrections.
Fix the rule, not the sentence
At FrodX, we call this system Content Factory. My voice is written down in a style guide with more than thirty banned phrases and expressions. A draft has to score 9.2 out of 10 before I see it, and then GPT and Gemini critique it. Every correction I make becomes a rule for every text that follows.
The best example is the em dash. Pew Research Center found that it now appears almost twice as often on the English-language web as in 2023, from 5.8 to 11.2 per 10,000 words. Pew counts it among the signs that give away AI-written text, so we have banned it. Readers can’t say why a text sounds natural. They can tell when it doesn’t.
I was slower at first
At first, the factory made me slower than I had been without it. I wrote prompts, fixed sentences, reworked the Croatian versions by hand, and always found something else to change.
Progress came when I stopped making corrections in the text and started making them in the rules. We also handed Croatian transcreation over to an OpenAI model that writes better Croatian.
Today I no longer rewrite drafts or write prompts. A column comes out of a conversation, the way you would brainstorm with a colleague. The conversation behind this column started with: “I want to write a column. Ready?” Then the machine and I looked for the thesis, checked the numbers, argued about HubSpot and threw out a long list of headlines. The machine wrote the draft, and I decided what it claims. Behind that conversation are months of building rules.
Why not just use Marketing Studio?
A fair question: why build your own factory when HubSpot has Marketing Studio? Its content agent writes in a brand voice that HubSpot extracts from samples of your writing.
For many companies, that is enough. If you write in one language, publish in HubSpot and speak with one voice, you can adapt to the tool, so go with Marketing Studio. A custom-built factory starts to pay off when you write in several languages, publish elsewhere too, draw on your own knowledge base and have your experts approve the copy. At that point, the tool has to adapt to you.
Either way, the answer is uncomfortable for HubSpot and for us. A sample shows the machine how you have written. Only someone who knows can tell it what you must never write. If you have rules, any factory works, HubSpot’s included. If you don’t, none does, ours included.
You won’t learn agentic writing in a prompting course. Open the last three machine drafts you corrected. If you fixed the same thing in all three, you repeated a rule three times without ever writing it down. Write it down. Your taste will keep working even when you are not there.
The word from October no longer appears in our texts. Nobody writes it anymore.
igor.pauletic@frodx.com
P.S. Did you spot the machine before I admitted it?
See how HubSpot connects content marketing, CRM and sales in one process.
