conorbronsdon__avoid-ai-writing/detector/CATEGORIES.md

Category map: references/patterns.md ↔ detector

This table is the anti-drift contract between the human-readable rules in ../references/patterns.md and the executable engine in patterns.js. When you add a rule to the skill, decide here whether it's regex-detectable (give it a detector type) or LLM-only judgment (mark it so). When you add a detector type, point it back at the skill section it enforces.

The engine exposes 53 issue types (see TYPE_LABELS in patterns.js). The skill has more ### sections than that — the gap is not missing coverage, it's rules that are judgment calls a regex can't make. The three groups below account for every entry on both sides.

Three counts coexist on purpose and should not be forced to match: the README's pattern-category count (the human-facing prose catalog, derived from references/patterns.md and guarded in CI), the engine's 53 types (which split the vocabulary tiers and add stylometric signals), and references/patterns.md's ### sections (which also include writer-side tests with no detectable form). The categories.test.js enforces the engine ↔ this-file mapping, and checks every prose statement of the engine type total against TYPE_LABELS.

A. Direct mapping (skill rule → detector type)

Detector type	Label	references/patterns.md section
tier1 / tier2 / tier3	AI vocabulary / Word cluster / Overused word	Words and phrases to replace (load-bearing: immediate abstract-noun allowlist only; literal, predicative, and unlisted forms pass; technical context mode suppresses eight technical-legitimate terms: robust, comprehensive, seamless, ecosystem, leverage, facilitate, underpin, streamline)
tier1-clarity	Wordiness	Words and phrases to replace (Tier 1B)
transition	AI transition	Transition phrases to remove or rewrite
template-phrase	Template phrase	Template phrases (avoid)
tier3-phrase / tier3-phrase-cluster	Boilerplate phrase / cluster	Template phrases (avoid)
chatbot	Chatbot artifact	Chatbot artifacts
sycophantic	Sycophantic tone	Sycophantic tone
filler	Filler phrase	Filler phrases
hollow-intensifier	Hollow intensifier	Filler phrases (intensifiers), except context-dependent actually (see §C)
generic-conclusion	Generic conclusion	Generic conclusions
social-cta-closer	Engagement-bait closer	Social endorsement closers
future-narrative	Generic future narrative	Generic future-narrative closers
lets-construction	"Let's" opener	"Let's" constructions
reasoning-artifact	Reasoning artifact	Reasoning chain artifacts
significance-inflation	Significance inflation	Significance inflation
novelty-inflation	Novelty inflation	Novelty inflation (the invented-concept-labels sub-rule is LLM-judgment only — open-ended coinages aren't regex-matchable)
real-actual-inflation	"Real/actual" inflation	"Real/actual" adjective inflation
vague-attribution	Vague attribution	Vague attributions
emotional-flatline	Stock reaction framing	Stock reaction framing / Superficial -ing analyses (stable API type; style-only for authorship scoring under the precision-first evidence policy)
lingering-attention	Lingering-attention claim	Lingering-attention claims (noun-anchored frames only — the bare "I keep coming back to X" stays LLM-judgment, since a following reason clause makes it legitimate and isn't regex-detectable)
cutoff-disclaimer	Cutoff disclaimer	Cutoff disclaimers
false-concession	False concession	False concession structure
rhetorical-question	Rhetorical question	Rhetorical question openers
formulaic-opener	Formulaic opener	Formulaic challenges
speculative-opener	Speculative scenario opener	Speculative scenario openers
confidence-calibration	Confidence stacking	Confidence calibration phrases
hedge-stack	Hedge-stacked prediction	Hedge-stacked predictions
parenthetical-hedge	Parenthetical hedge	Parenthetical hedging
hashtag-stuff	Hashtag stuffing	Hashtag stuffing
bullet-np-list	Bullet-NP list	Bullet lists of bare noun phrases
title-case-header	Title Case header	Title case headings
em-dash / formatting	Em dash overuse / Formatting	Formatting
uniformity	Rhythm uniformity	Rhythm and uniformity
low-ttr	Low vocabulary diversity	Vocabulary diversity (stylometric)
ai-placeholder	Unfilled placeholder	Unfilled placeholders
ai-citation-markup	Chatbot citation markup leak	Chatbot citation markup leaks
ai-utm-source	AI-tool URL parameter	AI-tool URL parameters
smart-punct-signature	Smart-punct signature	Formatting (curly quotation marks) — partial
unnecessary-hyphenation	Unnecessary hyphenation	Unnecessary hyphenation (curated open, closed, and position-dependent subclasses only)
launch-intro	Launch-copy introduction	Launch-copy dramatic introductions ("Meet X," plus one of "your new favorite", "your new go-to", "the new home of", "the new way to", "the new standard in/for", or a bare home/way/standard at end of clause; and "Think X meets Y". X is ONE capitalized 2-30 character token, so two-token names miss. Bare "Enter X.", bare "Meet X, your new [role]" and "Say hello to X" stay LLM-judgment; person-name variants are disclosed residue — see the entry)
crowd-contrast	Dramatized crowd contrast	Dramatized contrast against the crowd (three branches: "was|were|is|are" + "still" + a dismissive verb in -ing, plus the stereotyped "writing|wrote think-pieces" and "play|plays|played|playing catch-up" wordings, which need no marker; the crowd is the closed list "everyone else|others|the industry|the market|the competition", so "every competitor" and "our rivals" miss — ordinary simultaneity without those verbs stays clean; every literal progressive use of them is accepted, disclosed residue, and so are literal "wrote think-pieces" / "played catch-up" contrasts — see the entry)
fake-casual-prop	Fake-casual prop	Fake-casual register (six asterisk stage directions and the four (yes|no) x (really|seriously) parentheticals, nothing else. "chef's kiss" requires its apostrophe, and neighbours like "checks calendar" and "(yes, honestly)" are disclosed misses. Verdict closers, label-prefix openers, the self-QA volley and "because of course it does" stay LLM-judgment)
performed-insight	Performed-insight phrase	Performed-insight phrases — partial; literal-sense exclusions documented in references/patterns.md
negation-chain	Negation chain	Negation chains — partial; three-item deterministic threshold documented in references/patterns.md
dev-blog-boilerplate	Dev-blog boilerplate	Dev-blog boilerplate — partial; literal-sense exclusion documented in references/patterns.md

Partial map: smart-punct-signature fires only when curly quotes co-occur with an em-dash, an Oxford comma, and clean typing (≥80 words) — never on curly punctuation alone. The references/patterns.md Formatting rule treats curly quotes as a weak, corroborating signal in plain-text contexts and excludes apostrophes. The two agree in spirit (curly punctuation is never conclusive on its own) but differ in mechanism — so this is a partial map, not 1:1. performed-insight omits literal "the punchline" and "worth naming" senses; dev-blog-boilerplate omits literal "batteries included." Those ambiguous surface forms remain judgment rules rather than deterministic matches. negation-chain requires three short sentence-initial "no …" items; two-item chains remain judgment calls.

B. Detector-only (stylometric / fingerprint — no skill prose)

These extend the skill with signals that work as math over the whole document, not as a phrase a human editor would look up:

Detector type	Label	Why it's engine-only
punct-distribution	Punctuation distribution	Per-paragraph punctuation uniformity
fnword-trigram-entropy	Grammar repetition	Function-word trigram entropy
cross-para-burstiness	Cross-paragraph rhythm	Sentence-length variance across paragraphs
normalization-flag	Bypass-trick chars	Zero-width / homoglyph humanizer-bypass detection

C. Skill-only (LLM judgment — no detector type)

Rules that require reading for meaning, so they live in the skill prose and are applied by the model, not the regex engine. Listed so future contributors don't mistake their absence for a coverage gap:

• Transformation crutch (P2 unexplained relabeling, evaluated across the passage; literal and explained changes pass)
• False agency (P2 obscured accountable decision-makers; conventional personification, collective actors, emergent change, and causal dependence pass)
• Audience-fit domain-term collision (ambiguous generic proof/proof point in cryptography; separate from vocabulary tiers)
• Consequence-free "This matters because" / "here's why that matters" restatements (concrete consequences pass)
• Repeated empty concession pairs under Manufactured punchlines (P2; meaningful concessions and isolated intentional pairs pass)
• Synonym cycling
• Copula avoidance
• Promotional language
• Context-dependent actually as a hollow intensifier (delete it when it only adds emphasis; keep it when it carries a named correction or expectation gap). The same token performs both jobs, so matching it unconditionally would flag ordinary corrective prose.
• Sentence structure: "It's not X — it's Y" / split-sentence form / multi-negation countdown / tailing negation
• Structural issues / Excessive structure / Inline-header lists / Numbered list inflation
• Moral-adjective category errors (including ontological slop on assumptions, gratuitous universal quantifiers)
• Invented contrast-pair mirroring
• Hyphenated modifier stacking
• Unnecessary hyphenation outside the detector's curated subclasses
• False ranges
• Notability name-dropping
• Vague third-party validation
• Self-labeling significance
• Wall-of-text replies (missing line breaks) (tried as a detector — "reply-length, >=4 sentences, zero newlines" — and reverted; it fires on any ordinary short paragraph, not just conversational-reply register, so it stayed judgment-only. See the NOTE in patterns.js near the bullet-NP-list block)
• Recap-flattery opener
• Acknowledgment loops (tried as a detector — "you're asking about", "the question of whether", "to answer your question" — and retired (#239): the first opens ordinary replies and support answers ("You're asking about the retry limit. It is five by default..."), the third opens replies to a named earlier question ("To answer your question from Tuesday: the invoice went out on the 3rd..."), and the second is standard academic and analytical English. Both reply openers are document-initial in those examples, so position cannot separate them from the tell. Judging it needs reading whether the restatement adds anything before the answer arrives)
• Narrated candor (tried as a detector and reverted: the phrasings are shared with idiomatic conflict-of-interest disclosure ("in the interest of full disclosure, I own shares in...") and with the ordinary English comparative ("I'd rather die than let you drive"), so any regex tight enough to avoid those stopped matching the tell. Judging it needs reading whether the clause carries information or only announces that information is coming)
• Immaculate typography in casual registers (folded into the Formatting section — same weak-signal tier as curly quotes, not a standalone category)
• Subjectless fragments and agentless passives (docs and changelog registers are carve-outs — the fragment is the correct form there)
• Diff-anchored writing (changelogs, release notes, and migration guides are carve-outs)
• Manufactured punchlines / staccato drama (including repeated setup/reversal punchlines: P2 only when repetition replaces concrete claims; supported reversals and intentional comedy, fiction, speeches, and quotations pass)
• Aphorism formulas (a regex for "X is the Y of Z" would flag ordinary genitive copulas — "Paris is the capital of France")
• Stacked rhetorical questions (interviews, FAQs, and dialogue stack questions legitimately; a regex can't read register)
• Same-opener sentence runs (whether the anaphora is earned is the whole judgment; pronoun-opener runs are ordinary narration)
• Stranded auxiliary contrast (a single instance is legitimate style; only density across a piece distinguishes voice from tic)
• Colon into a triple (three-item lists are often simply true, especially in technical writing — weigh by genre, not per hit)
• When to rewrite from scratch vs. patch
• Severity tiers (P0 / P1 / P2)
• Self-reference escape hatch
• Output format

Partial: the skill's six context profiles map to the engine's broader options.contextMode values as documented in references/patterns.md#detector-mode-mapping. Only technical currently changes flagging behavior (skips title-case headers and suppresses eight technical-legitimate terms: robust, comprehensive, seamless, ecosystem, leverage, facilitate, underpin, streamline); marketing and personal are accepted and reported in stats but score like general. Full profile-based tolerance remains an LLM-side judgment.