Prose Quality

scripts/check-copy/rules/prose.mjs is the actual gate — this skill explains what it checks and why. Where tells-detector catches vocabulary and phrasing that specifically mark text as LLM-generated, this rule set catches prose weaknesses that make writing hard to read regardless of whether an AI wrote it — the classic style-guidance lineage (Orwell's six rules, Strunk & White's "omit needless words" and active-voice preference, Gary Provost's sentence-rhythm principle, Helen Sword's "zombie nouns") made mechanical.

Backed by real libraries rather than hand-rolled heuristics: the retext ecosystem (retext-passive, retext-intensify, retext-repeated-words) for grammatical pattern detection, write-good for wordy-phrase and cliché heuristics, and flesch / flesch-kincaid for genuine readability formula scores — the same class of tool proselint and Vale use, not a from-scratch reimplementation of syllable counting or sentence tokenization.

What each rule catches

• prose/passive-voice — flagged by retext-passive's grammatical pattern match. Orwell's rule 4 and Strunk & White's Rule 14: active voice names an actor, passive voice often hides one ("mistakes were made" vs. "we made mistakes").
• prose/weasel-word — vague qualifiers ("very," "quite," "arguably," relative "that") from retext-intensify. These words often do no real work; cutting them rarely changes the sentence's truth value.
• prose/repeated-word — accidental lexical duplication ("the the mat") from retext-repeated-words. Always an error, never a style choice.
• prose/hard-to-read-sentence — individual sentences flagged by retext-readability as dense relative to a general-audience target.
• prose/write-good — wordy or unneeded phrases ("due to the fact that" → "because") and clichés, from the write-good library. Runs alongside retext-passive with overlap deduped so a single passive clause isn't reported twice from two sources.
• prose/reading-ease / prose/grade-level — document-level Flesch Reading Ease and Flesch-Kincaid Grade scores, computed from real sentence/word/syllable counts (not a per-sentence proxy). The default contract targets general-audience copy: Reading Ease floor of 50, Grade ceiling of 10. A technical document that must use jargon will legitimately score below this — see the note on extending the contract below.
• prose/low-burstiness — sentence-length standard deviation below 3 words across a document with 4+ sentences. Gary Provost's principle: uniform sentence length reads as mechanical regardless of word choice; human writing mixes short and long on purpose.
• prose/nominalization-density — words ending in -tion/-ment /-ance/-ence at more than 3% of total words. Helen Sword's "zombie nouns": turning a verb into an abstract noun ("implementation" for "implement") removes the actor and adds length without adding meaning.
• prose/wall-of-text — a single paragraph over 100 words. A wall of text reads as dense and mechanical regardless of medium — this is the universal counterpart to the landing-page format's tighter 60-word mobile ceiling (format/wall-of-text in rules/formats/landing-page.mjs), and applies to every draft the gate checks, not just that one format.

Enumeration reads better as a list

No rule catches this mechanically, but it's the same family of defect as low-burstiness and nominalization density: a shape that reads as mechanical or dense even though no single sentence is wrong. A paragraph naming two or more things in a row ("...went out at 457 characters... A Meta ad shipped with...") reads denser and more repetitive than the same content broken into a list or numbered points, especially once a reader is scanning rather than reading straight through. The gate won't fail a paragraph for this; rewrite toward the list anyway before presenting a draft.

Errors vs. warnings

Only prose/repeated-word is error-severity — an unambiguous defect. Everything else is a warning: real signal, but a judgment call against the brief's register, per hedgehog-copywriting-loop's step 3. A warning-only report still passes the gate; don't chase every warning to zero at the cost of a sentence that reads worse for it.

When a violation looks wrong

A domain that genuinely requires a lower reading-ease score (deep technical documentation, a regulatory notice) or heavier passive voice (a process description with no clear actor) is a signal the brief's register doesn't match this core's default general-audience contract — raise it explicitly rather than silently accepting a failing or warning-heavy report as "good enough for this case."