Most people set up an AI writer the same way. They write down the words the agent is not allowed to use, paste the list in, and hope for the best. It catches the obvious cases. "Delve" goes. "Unlock the power of" goes. "In today's fast-paced world" goes. And then the writer produces a post that avoids every banned phrase and still sounds nothing like you.
A banned words list is a filter. It removes. It cannot add. And the thing that makes a post sound like your brand is the presence of certain moves, not the absence of a few phrases. This post is about the other half: the positive rules you can layer on top so a draft arrives sounding like you the first time round, rather than needing the editor agent to drag it there.
Why the banned list stops working
Banned words are easy to write and easy to reason about. That is why every guide starts there. But the words themselves are a symptom. The reason "delve into" grates is that the sentence around it is padded, vague and reaching for weight it has not earned. Strike the phrase, and the sentence underneath is still padded, vague and reaching. You get a draft that is technically compliant and still generic.
The other trouble is scale. A useful banned list grows fast. Ten becomes thirty, thirty becomes a hundred, and now you are asking the writer agent to keep a small dictionary in its head while it writes. It will drop things. And the moment a new AI cliche shows up, your list is out of date.
Positive rules do not have that problem. "Write in the second person" covers every sentence in the post at once. It does not need updating when a new filler phrase enters the vocabulary. It shapes the draft rather than trimming it.
The four kinds of positive rule worth adding
There are four places where a writing rule pulls its weight. Adding one rule from each of these buckets does more for a draft than another twenty banned phrases.
Sentence length
Prose sounds different when the sentences do. Long sentences read like a report. Short sentences read like a person talking. A brand that wants to sound approachable and a brand that wants to sound authoritative are not choosing different vocabularies as much as different pacing.
A rule that pins this down might read: "Keep sentences short. One idea per sentence. Break a long one in two." That is three sentences of guidance, and every paragraph the writer produces has to answer to them. The result is not a post full of choppy fragments. It is a post where a long sentence has to earn its length, because the default is short.
Point of view
Point of view sets the whole relationship between the post and the reader. Second person ("you point RankMill at your site") sounds like advice. Third person ("users point the platform at their sites") sounds like a manual. First person plural ("we point RankMill at your site") sounds like a pitch. Pick one and hold it.
The rule is simple: "Write in the second person." One sentence, and it decides the pronoun of every explanation in the post. Without it, the writer agent will drift between the three, sometimes inside the same paragraph.
Evidence style
Every brand has a stance on numbers, quotes and named examples. Some want case studies and named customers in every post. Some want none of it. Most want the truth, which is: use numbers when you have them and never make one up.
A rule that captures the strict version: "No statistics, percentages, case studies or customer quotes unless they were given to you as facts. If you do not know a number, write the sentence without one." That closes the door on the writer agent guessing a plausible-sounding figure to fill a paragraph. It also closes the door on the invented customer testimonial, which is the worst kind of drift because it looks like proof.
Forbidden rhetorical moves
This is the one people miss. Certain rhetorical moves are not phrases you can ban, they are shapes the writer keeps reaching for. The throat-clearing intro. The "in a world where" opener. The closing paragraph that summarises what you just said. The teasing subhead that hides the answer.
You cannot ban these by word. You have to ban the move. A rule like "Lead with the useful part. Do not open with a paragraph about how the topic matters. Start with the thing the reader came for" tells the writer agent to skip a specific structural habit, not a specific phrase.
A worked example, in the shape of a writing rules block
Writing rules in a RankMill project are a plain list of sentences. Every agent working on the project reads them and has to follow them. Here is what a set built for a small B2B software company might look like, once the four buckets above are covered.
Write in the second person. Address the reader as "you", never as "the reader" or "users".
Keep sentences short. One idea per sentence. Break a long one in two.
Never use an em dash. Use a comma, a colon or a second sentence instead.
No statistics, percentages, case studies or customer quotes unless they were given to you as facts. If you do not know a number, write the sentence without one.
Lead with the useful part. Do not open with a paragraph about how the topic matters. Start with the thing the reader came for.
Do not end the post with a section that summarises what came before. Stop when the last useful point is made.
Headings are sentence case. They say what the section is about rather than teasing it.
British spelling: colour, behaviour, organise, recognise.
That is eight rules. It is not a long list. Every one of them shapes the whole draft.
Notice what is not in there. There is no banned phrases list at all. Not because banned phrases are useless, they still have a place, but because with the positive rules doing their job, the phrases the writer would reach for do not come up in the first place. A writer trying to keep sentences short and skip throat-clearing intros is not writing "in today's fast-paced world" anyway.
Before and after on the same topic
Here is the opening the writer agent produces for a post titled "How to plan a content calendar", running against a project with only a banned words list.
Planning a content calendar is one of the most important tasks a marketing team can take on. In a world where audiences expect a steady stream of quality content, having a plan is what separates the brands that grow from the ones that stall. This guide will walk you through the essential steps for building a calendar that works for your team. Whether you are a solo founder or part of a larger organisation, the principles are the same.
Nothing in that opening breaks a banned words rule. There is no "delve", no "unlock", no "fast-paced world" (it went with "in a world" instead). And it says almost nothing. It takes four sentences to reach the promise that the post will explain something.
Here is the same writer, on the same topic, in a project with the rules above.
A content calendar is a list of what is going out and when. To plan one, you need three things: the posts you already have, the slots you want to fill, and a rule for what goes where. Start with the slots.
Four sentences. Three of them are shorter than the ones above. The reader is at the first useful instruction by sentence four, instead of somewhere in paragraph three. Nothing has been trimmed after the fact. The draft came in that way.
The difference is not the banned list. Both projects had one. The difference is that the second project told the writer how to think about a sentence, a point of view and an opening, and the first project only told it what words to avoid.
What the editor agent is left with
The editor agent is not there to rewrite the whole post. It is there to catch the things the writer got wrong on a first pass. If the writer is producing drafts that already keep sentences short, hold the second person and skip the throat-clearing intro, the editor's job shrinks to the actual editing: a claim that needs sourcing, a paragraph that is out of order, a phrase that reads oddly in context.
The alternative is the setup you started with, where the writer produces a generic draft that dodges the banned words, and the editor agent spends every turn trying to make it sound like the brand. That works. It is slow and it burns runs. And you can see the seams, because a post edited into voice reads differently from a post written in voice.
Positive rules move that work back one step. By the time the editor sees the draft, the voice is already there.
Where to start if you have not done this
If your project has only a banned words list, add three rules this week. Pick point of view, pick a sentence-length stance, and pick one forbidden rhetorical move that the writer agent keeps reaching for. Write each one as a plain sentence, drop it in, and generate a post. You will see the shift on the first draft.
Then, over the next few posts, add the ones that keep coming up in your own edits. A rule earns its place when it stops a mistake you were fixing by hand. That is the whole test.