The editor agent in RankMill takes any change you can describe in words. That is the good news and the trap. Ask for something vague and you get a vague answer: a shuffled adjective here, a synonym there, a paragraph that reads almost the same. Ask for something specific and you get the revision you actually wanted, delivered as a diff you can read line by line.
This post is a working taxonomy of edit prompts, sorted by what they change. Each section has a weak version and a specific one, applied to the same paragraph, so you can see the shape of the difference. Use it as a checklist the next time a post is close but not right.
Why vague prompts produce cosmetic edits
The editor agent works from three things: the post as it stands, the conversation so far, and the writing rules and website profile for the project. When your instruction is thin, the agent has nothing to anchor a real change to. It falls back on the safest thing it can do, which is rewrite the surface of the sentence.
"Make it punchier" is the worst offender. There are a dozen ways to make prose punchier, and most of them contradict each other. Cut adjectives. Break long sentences. Front-load verbs. Delete the second half of every paragraph. The agent picks one, applies it lightly, and hands you back something that reads mostly the same because it had to guess what you meant.
The fix is not to write longer prompts. It is to name the thing you want changed.
Four things you can change
Every useful edit falls into one of four categories: structure, argument, evidence, or tone. If you know which one you are asking for, the prompt writes itself.
Structure: what goes where
Structural edits move, split, merge, cut or add. They do not touch the meaning of any individual sentence. They rearrange.
Weak: "This section feels off."
Specific: "Move the paragraph starting 'Most teams get this wrong' up to become the second paragraph, right after the opener. Then cut the current second paragraph, it is repeating the same point."
The weak version leaves the agent guessing what "off" means. The specific one names the block, names the destination and names the cut. The diff comes back showing exactly those three moves.
Other structural prompts that work:
- "Split the section 'Setting up a project' into two, one on adding the URL and one on writing rules."
- "Add an H3 called 'Common mistakes' at the end of the main section, three short bullets, no intro sentence."
- "The bullet list under 'What to check first' should be a numbered list, the steps are ordered."
- "Delete the whole 'Why this matters' section, the reader already knows."
Notice how each one names the target and the action. The agent does not have to interpret.
Argument: what the post is claiming
Argument edits change what the post says, not just how it says it. This is where vague prompts hurt most, because the agent will not rewrite a claim unless you tell it to.
Weak: "Make the case stronger."
Specific: "The section on scheduling says it is optional. It is not optional for anyone running more than one project. Rewrite it so the default position is that you turn it on, and the caveat is when you might not."
The weak version gives the agent nothing to work with, so it adds an adverb or two. The specific one flips the argument's polarity and tells the agent what the new default position should be. The diff shows a real rewrite.
More prompts in this shape:
- "The opening admits the tool has limits before saying what it does. Swap the order. Lead with what it does."
- "This paragraph gives three reasons and then walks back the third one. Cut the third reason, do not defend the other two."
- "The conclusion says 'it depends'. It does not depend. Say the answer, which is to turn it on for teams larger than one."
- "You are comparing us to a competitor and being polite about it. Do not be polite. Name the specific thing they cannot do that we can, which is opening a pull request on the writer's own repo."
Argument prompts work best when you state the new position you want, not just that the old one is weak.
Evidence: what backs the claims up
Evidence edits change the examples, the analogies, the concrete details. They are the ones most likely to sound cosmetic if you are not careful, because swapping an example is a small textual change even when the argument shifts under it.
Weak: "Add an example."
Specific: "The paragraph about writing rules has no example of a rule. Add one, using our own 'Never use an em dash' rule, and show it as a bullet inside the paragraph so the reader sees what a rule looks like in practice."
The weak version gets you a generic example that could apply to anything. The specific version names the paragraph, names the example, and describes the format the example should take.
Others in this family:
- "The abstract paragraph about voice needs a before-and-after: one sentence in a brochure voice and the same sentence in ours."
- "Cut the invented statistic in the second paragraph. Replace it with a specific behaviour you can point to, like the website profile being editable as a diff."
- "The section on team scale keeps saying 'small teams'. Give the actual sizes we mean, which are solo founders, teams of one to five, and small agencies."
- "The claim about publishing needs the mechanism spelled out: pull request opens, you merge, the post counts as published on merge."
Evidence prompts are strongest when they name the thing the reader should be able to picture after the edit.
Tone: how it sounds
Tone edits change the register, the rhythm and the vocabulary without changing what the post claims. They are the most subjective category, so they are also the ones where "make it punchier" lives. Do not ask for a mood, ask for a mechanism.
Weak: "Make it warmer."
Specific: "The whole post is in the third person, 'users' and 'teams'. Rewrite in the second person throughout, 'you' and 'your'. Keep every claim the same."
The weak version leaves the agent to guess what "warmer" is. The specific version names the mechanism, the second person, and the scope. The diff shows a global rewrite of pronouns, nothing else changed.
More tone prompts that reliably land:
- "Split any sentence longer than twenty words. One idea per sentence."
- "Cut every adverb ending in -ly. If the verb needs help, pick a better verb."
- "The section headings are teasers. Rewrite them so each heading says what the section is about."
- "There are three sentences starting with 'It is important to note'. Delete that opening in all three, keep the rest."
Tone is the category where writing rules do the heavy lifting for you. If a rule already bans a phrase, do not repeat the rule in a prompt. Point at where the post breaks it.
Side by side on the same paragraph
Take one paragraph and run each category over it. Start with this original:
RankMill is a really powerful platform that helps teams unlock the power of AI to create amazing content. Users can leverage cutting-edge agents to streamline their workflow and produce high-quality blog posts at scale.
A structural prompt: "Split this into two sentences, one on what RankMill is and one on what teams do with it." The diff cuts the sentence in half but leaves the words alone.
An argument prompt: "This says the platform helps teams produce at scale. Scale is not our pitch, control is. Rewrite the second sentence so the point is that you approve every step, not that you produce more."
An evidence prompt: "Delete 'high-quality' and 'amazing', they mean nothing. Replace the vague talk about agents with the actual four: research, writer, editor, and the human approval gate."
A tone prompt: "Cut every filler word. No 'really', no 'unlock the power of', no 'cutting-edge', no 'leverage'. Second person throughout. Two sentences, both short."
Each prompt changes a different part of the same paragraph. You can stack them across turns, and because the editor agent keeps the conversation, you can say "now do the tone one on top of the argument version" and it will.
Working the conversation
A few habits pay off across every edit.
One change per turn. The diff is easier to read and easier to reject if one thing changed. Save the multi-part rewrites for cases where the parts genuinely depend on each other.
Name the target. A section, a paragraph, a sentence, a phrase. "The opening", "the bit about pricing", "the third bullet under 'Getting started'". The agent cannot edit what you have not pointed at.
State the new state, not just the problem. "This is too long" is a complaint. "Cut this to three sentences" is an instruction.
Clear the conversation when you change direction. Editor memory is useful when you are refining one thread. It gets in the way when you have moved on. Start fresh and the next diff will not be shaped by an argument you already dropped.
Read the diff. Every edit turn comes back as a diff on the post, not as a fresh copy. Skim the removed lines as well as the added ones. Sometimes the agent cut something you wanted to keep, and the fix is one more prompt saying so.
The editor agent will do the work if you tell it what work to do. Vague in, vague out. Specific in, and you get the revision you actually asked for.