The Paragraph Rule

"Paragraphs run 3–5 sentences, hard max — the finished report must scan at first glance."

That line isn't writing advice. It's quoted word for word from the contract our AI writes strategy reports under: the set of instructions handed to the machine every time it writes, governing what it is and isn't allowed to put on a page.

No engineer came up with it. Our editor set it: a human rule about how finished work must read. On July 16 it went in as law, written into the system by midnight the same day, before the wave of reports it governs had generated a single specimen.

I want to walk you through this one small rule, because it turned out to be the highest-leverage edit in our whole report system, and because the reason it works teaches you something useful about every piece of AI-written material that crosses your desk.

Density is where machine verbosity hides

If you've read much AI-generated writing (and if you run a brand, you're reading more of it than you think), you've felt a specific kind of fatigue without necessarily naming its cause. The sentences are fine. The grammar is clean. And somewhere in paragraph three your eyes slide off the page.

The cause is density. Left to its own defaults, a model doesn't write paragraphs the way a person does — it accumulates. Clause after qualifying clause, sixth and seventh sentences that restate the fourth, every point cushioned and elaborated until the paragraph is a wall. Each sentence individually passes inspection. The block as a whole is unreadable. Our editorial standard calls it what it is: "Density is a defect."

This matters for a strategy report more than for almost any other document, because of who reads one and how. A strategy report gets read by an operator flipping pages, someone deciding in seconds whether a section deserves minutes. If it doesn't scan at first glance, the thinking inside never gets a hearing.

The rule follows the reader, and we know that because we write the opposite kind of prose on purpose in another room of the house. The content our answer-engine tool produces is written for machines to quote, because that is what ChatGPT and Perplexity lift cleanly. Our standard for strategy reports names that style and then deliberately walks away from it. A report is judged by a person in long form.

Same company, two opposite sets of rules, and each one is correct for the reader it serves.

So when you're evaluating AI-written work (a vendor's report, an agency deliverable, your own team's output), density is one of the fastest tells there is: not whether it reads as human, but whether it scans. Machine fluency is cheap now; machine restraint isn't, because restraint has to be imposed.

How a human rule becomes machine law

This is what imposing it looks like inside our system, step by step.

The contract states the rule plainly, and then tells the model what to do at the boundary: "If a paragraph needs a sixth sentence, split it into two prose blocks." That's the instruction side, the rule as the model receives it.

The paragraph rule doesn't work alone, though. It sits inside a set of instructions that are all aimed at the same reader flipping pages.

Short paragraphs are the floor of that design. The headers do the signposting, the opening paragraph delivers the verdict, the anchors catch the eye, and the paragraphs underneath are short enough that none of it gets buried.

Then an automated check inspects the model's work, because an instruction without enforcement is a suggestion.

One detail in the counter carries a design principle. The counter is deliberately forgiving about counting: it knows "Dr." and "e.g." don't end sentences, so a report never dies on a miscount. It also knows what not to count. Section titles, list items, and the brand's own quoted words are all outside its scope; only flowing prose is measured, because only flowing prose is where density hides. But it is never forgiving about the rule. Precision about measurement, zero flexibility about the standard.

And the rule stacks on a prose-first mandate, which closes the obvious escape hatch. A model forbidden long paragraphs will happily hand you bullet soup instead, fragments masquerading as concision. The standard forecloses it: "flowing editorial paragraphs, just short ones." The standard is senior writing, not shredded writing. That mandate has its own mechanical check. A section where more than half the content is list items is what our standard calls "a field dump wearing a suit," and it fails the same way a nine-sentence wall does.

The page with the most riding on it gets no exemption. Every report opens with a one-page letter from the strategy team, and the letter's rules are the strictest in the document: flowing prose only, no subheadings, no bullets, no inline scores, under about 450 words, opening with a specific observation about this brand rather than a principle. The counter walks the letter's paragraphs too. The place a client reads first is held to the same paragraph law as the appendix.

And the person who bought the report keeps a lever of their own. Every document carries one structured revision: describe what is wrong and what right looks like, and the section is regenerated under the same contract. If that regeneration fails, the revision is refunded and the round is yours again. The law governs the machine; the judgment about whether the result reads right still ends with a human.

The lesson isn't about paragraphs

We've done far cleverer things to make machine prose read well. None of them bought as much as this one formatting constraint. The highest-leverage AI edit we've made is a formatting constraint a human refuses to bend.

The clever interventions all try to make the model want to write well, and the model doesn't want anything. It drifts back toward its defaults the moment the pressure lifts. The paragraph rule doesn't negotiate with the model's defaults at all. It defines a hard boundary on the output, checks it mechanically, and rejects work that crosses it. Taste, encoded as law, enforced without mood.

The fair objection is that this is a style preference dressed up as engineering, and the standard answers it by being honest about which of its rules are which. Sentence rhythm is in there too: a long evidential sentence followed by a short verdict "lands the way senior writing lands," while uniform medium-length sentences read as generated. That rule is real, and it is also judgment, so it is scored and flagged rather than enforced. Paragraph length is countable, so it became law. The line between taste and law is drawn exactly where a machine can measure without being wrong. Everything on the judgment side of that line stays with a person.

That's the transferable part, and it applies whether you're building AI systems or just buying their output. Your editorial standards — the ones a good human editor would enforce with a red pen — don't stop mattering because a machine is writing. They matter more, because the machine produces volume no human editor can hand-police. If you're evaluating an AI-powered vendor, this is a genuinely revealing question to ask: what human editorial rules govern your output, and what happens when the machine breaks one? A shop with a real answer has an editor in the loop. A shop without one is shipping the model's defaults.

Ours has an editor, and the arrangement is simple: the model brings the fluency, the editor brings the law, and the reports read like senior work because only one of those is negotiable.