Editorial Standards

Cleaning Up AI Prose:
An Editorial Rulebook
FOR AGENTS

A practical style guide for identifying and eliminating the patterns that make AI-drafted prose sound like AI-drafted prose. Six cadence antipatterns, six general principles, and a pre-publish checklist. Drop it into your system prompt and let your agent do the rest.
March 2026 · Brandon Huey

The single most recognizable tell of LLM-generated prose is staccato rhythm: short, punchy, declarative sentences arranged in contrasting pairs or rapid-fire lists. Every major language model produces this pattern by default, and it is immediately recognizable to editors, readers, and other LLMs. The pattern is so consistent across model families and providers that it functions as a stylistic fingerprint.

Most discussions of AI writing quality focus on factual accuracy. Hallucination rates, citation quality, and logical consistency receive the bulk of attention, and for good reason. But even when the facts are correct, the prose itself can undermine credibility by sounding like it was generated rather than written. This guide addresses that second problem: the cadence, rhythm, and structural patterns that mark prose as machine-produced, and the specific editorial rules that eliminate them.

These rules were developed through iterative use, applied across research briefs, technical documentation, and long-form analysis where LLMs performed the initial drafting. Each rule exists because the pattern it addresses appeared repeatedly, survived standard editing passes, and was identifiable by readers. The rules are prescriptive by design; they describe what to avoid and how to fix it, with concrete before-and-after examples.

Six patterns to detect and eliminate

2.1 False-Contrast Pairs
Two short sentences where the first negates and the second asserts. The contrast should land inside the sentence, not across a full stop.
Avoid
This is not a theoretical argument. It is standard financial engineering.
Use instead
This is standard financial engineering, not a theoretical exercise.
Avoid
It's not about speed. It's about survival.
Use instead
The issue is survival, not speed.
Fix: Merge into a single sentence using "rather than," "instead of," a comma, or a subordinate clause.
2.2 Sentence-Fragment Lists
Three or more ultra-short declarative sentences in rapid succession for rhythmic punch.
Avoid
Routers fail. Disks fail. TCP packets corrupt.
Use instead
Routers fail, disks fail, and TCP packets corrupt.
Avoid
The gap widens. The cost compounds. The window closes.
Use instead
The gap widens as costs compound and the window closes.
Fix: Join with commas, conjunctions, or subordination. If the items form a causal chain, connect them with "as," "while," or "and."
2.3 Dramatic One-Liner Closers
A single punchy sentence at the end of a paragraph or section, designed as a mic-drop.
Avoid
Permanent resistance is a terminal diagnosis.
Use instead
...no compensation package reverses the damage once permanent resistance has set in.
Fix: Absorb the closer into the preceding sentence. If the thought is important enough to keep, it is important enough to earn a full clause with context.
2.4 Parallel Structure Overuse
Back-to-back sentences with identical grammatical construction.
Avoid
Every dollar of cloud spend reduces EBITDA. Every dollar of GPU CapEx is invisible to EBITDA.
Use instead
Every dollar of cloud spend reduces EBITDA, whereas GPU CapEx is invisible to EBITDA entirely.
Fix: Vary the sentence structure. Use subordinate clauses ("whereas," "while," "rather than") to connect the two halves.
2.5 The Dramatic Pivot
A sentence that exists solely to signal a reversal or hidden insight, often starting with "But here's the thing," "The real question is," or using headlines like "The X Nobody Talks About."
Fix: Cut the throat-clearing and state the point directly. If the section heading needs a label, make it descriptive rather than theatrical.
2.6 Compressed Negation Tags
Appending "not X" to the end of a clause as a punchy kicker.
Avoid
Export controls bought time, not advantage.
Use instead
Export controls bought time without lasting advantage.
Avoid
The talent exodus is the leading indicator, not the lagging one.
Use instead
Talent flight precedes competitive decline, not the other way around.
Fix: Expand "not X" into a prepositional phrase ("without," "rather than") or a full clause.

Authority without assertion

Write as an informed analyst presenting evidence rather than a pundit delivering verdicts. The reader should trust the work because the reasoning is visible, because the sources are named, and because the data is specific enough to verify.

Never assert subjective rankings as objective fact. When a claim reflects someone's priorities or values, attribute it. If you cannot identify who holds the opinion, reconsider whether the sentence belongs in the piece.

Avoid
They failed at the thing that matters more.
Use instead
They failed at what policymakers said mattered more.
Avoid
The real risk is choosing nothing.
Use instead
Inaction carries more risk than a wrong choice, according to the firms we surveyed.

Six principles for connected writing

  1. Prefer connected phrases over sequential declarations. A paragraph should flow through an argument, not stack assertions. If you can hear a drumbeat when you read it aloud, rewrite it.
  2. Vary sentence length. Mix longer analytical sentences (25-40 words) with shorter ones (8-15 words). The problem is never a short sentence on its own; it is five short sentences in a row.
  3. Earn your emphasis. Bold, italics, and short sentences all serve the same function: emphasis. Use one at a time, sparingly.
  4. Lead with evidence, not verdict. Present the data or observation first, then the interpretation. The reader should be able to form their own conclusion before you state yours.
  5. Eliminate weasel words. This includes vague qualifiers ("some," "many," "significant," "arguably," "it is widely believed," "experts say") and throat-clearing preambles ("It is worth noting that," "It is important to understand that"). Either name the source, provide the number, or cut the qualifier entirely. If a claim cannot be backed with a citation or data point, the sentence is not ready to publish.
  6. Avoid the paradox formula. "X did the opposite of what it intended" is a valid observation stated once. Repeating the ironic-reversal framing across multiple sections makes the piece feel like it has one rhetorical move.

The em-dash trap

When LLMs are instructed to merge staccato sentences, they default to em dashes as the connector. Fixing six false-contrast pairs with six em dashes replaces one kind of monotony with another. The resulting prose has a distinctive look on the page: long sentences interrupted by paired dashes that the reader's eye learns to skip.

The fix is to vary connectors across the full punctuation inventory: periods, colons, semicolons, commas with conjunctions, parentheses, and subordinate clauses. No two adjacent sentences should use the same joining device. Em dashes are acceptable sparingly (one or two per page) but should never be the default merge tool.

Checklist

An audit in practice

The following is an unedited excerpt from an editorial audit of a published research brief. The article had already been fact-checked and was structurally sound, but running it against these rules surfaced seven violations across four rule categories. This is what applying the rulebook looks like in practice.

2.6 Compressed Negation Tags · 2 instances
Flagged
"The gap between per-token efficiency and aggregate capital efficiency is widening, not narrowing."
Fixed
"The gap between per-token efficiency and aggregate capital efficiency is widening."
Flagged
"value-stream efficiency, not units per hour, determined competitive survival"
Fixed
"value-stream efficiency determined competitive survival more than raw units per hour"
Pattern: Classic "X, not Y" kicker appended to a clause. The negation adds emphasis but contributes no information the reader does not already have from context.
2.2 + 2.4 Sentence-Fragment List · Parallel Structure
Flagged
"Airlines replaced Available Seat Miles with Revenue per Available Seat Mile. Telecom carriers replaced circuit-switched minutes with Average Revenue Per User. Manufacturing replaced units-per-hour with value-stream efficiency."
Fixed
"Airlines replaced Available Seat Miles with Revenue per Available Seat Mile; telecom carriers abandoned circuit-switched minutes in favor of Average Revenue Per User; and manufacturing shifted from units-per-hour to value-stream efficiency."
Pattern: Three consecutive sentences with identical "X replaced Y with Z" structure. Merged into a semicolon-separated series with varied verb choices ("replaced," "abandoned," "shifted") to break the rhythmic repetition.
2.3 Dramatic One-Liner Closer
Flagged
"The capital is flowing to throughput, not outcome."
Fixed
"The capital is flowing to throughput rather than outcome."
Pattern: Short punchy sentence at the end of an assessment section, functioning as a mic-drop. Also a compressed negation (2.6). Absorbed into a single clause with "rather than."
3.5 Weasel Words · 2 instances
Flagged
"aggregate cognitive output may actually be declining"
Fixed
"aggregate cognitive output shows signs of declining"
Flagged
"somewhere between a quarter and half of reasoning compute produces no measurable benefit"
Fixed
"27% to 51% of reasoning compute produces no measurable benefit"
Pattern: "May" hedges the central claim without adding nuance. "Somewhere between" is a vague rephrasing of the specific 27-51% data already cited in the same sentence. In both cases, the fix is to either commit to the claim with supporting evidence or use the specific numbers already available.
1 Unattributed Value Judgment
Flagged
"the industry's dominant metric, tokens per watt, is measuring the wrong thing"
Fixed
"the industry's dominant metric, tokens per watt, captures hardware performance without capturing whether the output is useful"
Pattern: "The wrong thing" is a value judgment presented as objective fact. The fix attributes the assessment to the evidence (Stanford's IPW research) and restates it as a specific, falsifiable claim rather than an editorial verdict.

Result: seven violations across four rule categories, all in prose that had already passed a fact-checking review. None of the fixes changed the meaning of the text; each preserved the original claim while eliminating the stylistic pattern that marked it as machine-generated. The seven fixes took under three minutes to apply.

Why this matters

LLM-assisted writing is becoming the default mode of content production across research, technical documentation, marketing, and journalism. The quality conversation has rightly focused on factual accuracy, but as hallucination rates decline (the best models now achieve sub-1% on standard benchmarks, according to the Vectara Hallucination Leaderboard), the remaining gap between AI-drafted and human-written prose is increasingly a matter of style rather than substance.

Prose cadence affects comprehension, trust, and retention in ways that go beyond cosmetic polish. A 2024 study by the Cornell Social Dynamics Lab found that readers rate AI-identified text as less credible even when the content is factually identical to human-written alternatives. For organizations publishing under their own byline, the gap between "factually correct" and "editorially credible" determines whether content actually influences its audience or merely occupies a page.

These rules exist to eliminate the artifacts that make AI's involvement obvious, not to eliminate AI from the writing process. The goal is prose that could have been written by a careful human author, regardless of whether it was.