The AI-style rulepack

writinglint-rulepack-ai-style is the first rulepack. recommended enables every rule below in confidence-aware auto mode, reporting medium- and high-confidence findings by default. warn. Rules are namespaced ai-style/<name>, so you override any of them in a config, e.g. 'ai-style/rule-of-three': 'off'.

The ★ structural rules match over the dependency graph — the ones a regex or POS-only linter cannot express. The rest are lexical / idiom / formatting rules.

Structure & cadence

Rule	Category	What it catches
★ corrective-antithesis	parallelism	The “X, not Y” staged contrast (“the graph, not the vibes”).
★ negative-parallelism	parallelism	“Not (only) X but (also) Y” — a signature LLM cadence.
★ copula-avoidance	significance	“stands/serves as a …” dressing up a plain “is a …”.
★ light-verb-role	significance	“plays a … role” — importance asserted, not shown.
★ participial-appendage	significance	Trailing “-ing” clause that editorialises the main clause.
★ rule-of-three	rule-of-three	Reflexive triads of coordinated adjectives or adverbs.
★ vague-attribution	vague	A bare, generic subject asserting a “that …” clause (“Studies suggest…”).
★ throat-clearing	meta	“It is important to note that …”. If it matters, just say it.
hedging-seesaw	balance	Relentless “While X… However, Y” balancing, density-gated — a position never taken.
★ comma-splice	rhythm	Two complete clauses stapled with a bare comma (“Thanks for the demo, I enjoyed it.”) — clipped parataxis performing breeziness.
★ agentless-opener	agency	“Notes attached, and they are …” — a verbless fragment that circles back with a pronoun instead of naming the doer.
★ setup-fragment	performance	“One thing I wanted to put on the table …” — a noun-rooted fragment staging a point instead of making it.

Vocabulary & idioms

Rule	Category	What it catches
ai-vocabulary	ai-vocab	Words LLMs over-use relative to human writers (delve, tapestry, …).
significance-idioms	significance	Fixed “inflated significance” idioms (rich tapestry, testament to …).
promo-idioms	promo	Travel-brochure / press-release idioms (nestled in the heart of …).
chatbot-idioms	meta	Editorialising / chatbot filler (“it’s worth noting…”).
opening-conjunction	conjunctions	Formulaic sentence-opening transitions (“Moreover,”, “Ultimately,”).
performed-candor	performance	Announced honesty (“to be fully transparent”, “I’ll be honest”) in place of the honest sentence itself.
★ filler-intensifiers	performance	Sincerity adverbs on first-person stance adjectives (“I am genuinely open”), plus density-gated intensifier spray.

Formatting tells

Rule	Category	What it catches
em-dash-overuse	formatting	Heavy em-dash use relative to sentence count (density-gated — a few em dashes never fire).
mixed-quotes	formatting	Straight and curly quotes mixed in one document — a paste seam.
generation-artifacts	formatting	Leftover chatbot citation tokens (oaicite, turn0search0, utm_source=chatgpt.com).
emoji	formatting	Decorative emoji in formal prose.

Why no Markdown or curly-quote rules?

Markdown syntax (**bold**, # headings) is not flagged: humans author Markdown all the time (READMEs, docs, notes), so it carries no AI signal on its own. Curly quotes alone aren’t flagged either — Word, Google Docs, and iOS auto-curl typed quotes; only mixing straight and curly styles carries signal. Writing casual prose where emoji are normal? Turn that one off: 'ai-style/emoji': 'off'.

The score is not a rule

Separately from the lints, the pack ships a stylometric score — a document-level 0–100 estimate of how AI-shaped the writing reads, from a small data-free classifier. It is deliberately not a rule: flags are local and structural; the score is global and statistical. See Consume as a library.