Ask anyone who has used an AI writing tool what went wrong and the answer is rarely "the grammar." It is "it didn't sound like us." The copy was fine and belonged to nobody.
RankMill attacks that problem from two directions. The website profile tells every agent what the business is: what it sells, who buys it, how it talks. Writing rules tell every agent how to write: the specific, checkable habits a draft has to follow. This post is about the second one, how it works today, and where its edges are.
Where the rules live
Writing rules belong to a project, not to your account or your team. A team running two very different sites, a developer tool and a bakery, does not want one set of rules for both.
You set them when you create the project. On the second step of the form, under Competitors and writing rules, there is a plain text box: one rule per line. Above it sits a row of suggestions you can click to add, and click again to remove:
- No em dashes
- No AI cliches
- Active voice
- Short paragraphs
- No fluff intros
- No emojis
- Second person
- Concrete examples
- Sentence case headings
- No exclamation marks
- Cite sources
- Practical takeaway
Each chip is a short label for a full sentence. "Short paragraphs" adds "Keep paragraphs to three sentences or fewer so the article stays scannable." We store the sentence, not the label, because the sentence is what the agent reads. A keyword like "short paragraphs" leaves the model to guess what short means. A sentence does not.
You can write your own rules alongside the suggestions, and you should. The suggestions are the generic habits most teams want. Your own rules are the ones that make a draft yours: "Call it a workspace, never an account." "Never promise a delivery date."
The business line
The same step asks one more question: What this business does. That sentence is saved as the first line of the project's rules, so the agents know the business in your words from the very first run. Leave it blank and the agent works it out from the site itself.
Who reads them
Every agent that produces or judges copy for a project gets that project's writing rules in its prompt:
- Research, while it writes the website profile
- The topic finder, while it proposes topics
- The writer, for every approved topic
- The editor, on every turn of an edit conversation
- The importer and the rewriter, when you bring in an existing post
- The profile editor, when you ask for an AI edit of the profile
- The auditor, which checks finished posts against them, and the agent that then fixes each finding
Nobody has to remember to paste the style guide into a prompt. A rule you add at creation is in front of every one of those agents on every run.
When the rules and the profile disagree
The profile and the rules usually agree, but not always. Research might describe a brand as playful and emoji-friendly after reading the homepage, while your rules say no emojis.
The shared instructions every RankMill agent receives settle this in one line. The website, its competitors, the writing rules and the profile together are the authority on voice, audience and terminology, and when two of them disagree, the writing rules win. The same instructions tell agents never to invent facts that are not in the material they were given.
Imports treat rules differently, on purpose
Importing an existing post has three modes, and the rules mean something different in each:
- Import as is stores your post verbatim. No agent runs, so no rule is applied.
- Restructure with AI cleans up the markup without changing a single sentence. Here the rules count only as formatting conventions, such as heading depth or list style. A rule about tone or vocabulary is ignored, because the author's words are already final.
- Rewrite it as our own keeps the facts and writes the post again in the project's voice, and here the rules and the profile are the authority.
If you import a post and it comes back with an em dash still in it, check the mode before you check the rules.
Rules for just one post
Sometimes one post needs extra handling: a piece that has to quote a named customer, or a launch post that must mention a specific date. When you Write from an idea, you can choose Change for this post. The dialog is explicit about what that does: rules set for one post are added to the project rules, not swapped for them. A post-level rule can add a constraint, but it cannot lift one the project already has.
What rules cannot do
Two limits are worth knowing before you rely on rules too heavily.
Rules are instructions, not a linter. They go into the prompt, and the model follows them. Nothing in RankMill rewrites the finished markdown to strip out a banned word afterwards. Models follow clear, specific rules well, but a slip is possible. That is why reading a post before it goes out still matters, and why the audit exists: it reads the project's completed posts, reports up to five of the most important problems, breaks from the project's rules included, then works through the fixes. The audit costs one run, and the fixes it makes afterwards are free.
Rules are set once. There is currently no screen for changing a project's writing rules after it exists, and the create form says so. Two things you can change at any time:
- The website profile. Its Voice & tone and Terminology & things to avoid sections are read by the same agents. Edit them by hand, or ask for an AI edit and review the diff before it applies.
- The edit conversation on a post. If a finished post breaks a rule, tell the editor. It revises the post and keeps its edits consistent with the writing rules.
Writing rules that work
Rules that change drafts tend to be:
- One behaviour per line. "Short paragraphs and no jargon" is two rules.
- Checkable. If two people could disagree on whether a draft follows it, rewrite it.
- Specific to you. The suggestion chips cover the generic habits. The rules that make posts sound like your company are the ones only you can write.
- Stated as what to do, with the alternative. "Never use em dashes. Use a comma, colon, or a second sentence instead" works better than "no em dashes," because it tells the writer what to reach for.
The short version
- Writing rules are per project, one rule per line, set on the create form
- Twelve suggestions are one click away, and your own rules sit beside them
- Every agent that writes, edits, imports or audits a post reads them
- When the profile and the rules disagree, the rules win
- An as-is import applies no rules, a restructure treats them as formatting only
- Rules are prompt instructions, so your own read and the audit are the backstop
- Rules cannot be edited after creation yet, but the profile and the edit conversation can
Voice is not something a model discovers by itself. It is something you write down, and writing rules are where it gets written.