Tells Detector

scripts/check-copy/rules/tells.mjs is the actual gate — this skill explains what it checks and why, so a tells/* violation gets fixed at the root cause rather than reworded just enough to dodge one regex. Never substitute a manual read of this list for actually running node scripts/check-copy/index.mjs; the whole point of this core is that the check is mechanical, not agent-graded.

Sourced from Wikipedia's "Signs of AI writing" essay (WikiProject AI Cleanup) and the AI-tell detection literature it's drawn from, plus prior art already proven out in the landing-page core's landing-copy skill. Frequency, not presence, is the real signal — one hedge or one long sentence in a document is normal prose. That's why vocabulary and phrase rules fire on every occurrence (a banned word has no legitimate frequency floor) while structural rules (em dash, rule-of-three) fire only past a per-document density threshold.

What each rule catches

• tells/banned-word — vocabulary that AI-detection research and frequency analysis (GPTZero's word-frequency data, Wikipedia's era-segmented list) show as wildly overrepresented in LLM output versus human baselines: delve, tapestry, robust, seamless, leverage, crucial, pivotal, elevate, unlock, and similar. These words aren't wrong in isolation — they're wrong because they're the default choice an LLM reaches for, so their presence at all is the tell.
• tells/banned-phrase — signposting fillers ("it's important to note that," "at its core"), vague attribution ("studies show," "experts agree" with no named source), and closing rituals ("in conclusion," "the future looks bright"). These are hedges and non-commitments, often traceable to RLHF training pressure toward inoffensive, non-committal text.
• tells/negation-formula — "it's not just X, it's Y" and its variants. State the positive claim directly instead of setting up a contrast that adds no information.
• tells/emphatic-fragment — a short standalone sentence built from absolute vocabulary ("Nothing else is left behind.", "Every time, not once.") that restates the previous claim instead of adding to it. The tell is the terse fragment used as emphasis, not short sentences in general. Cut it, or fold it into the sentence before it.
• tells/hedge-stack — pairing two hedges ("could potentially," "may eventually"). Pick one claim and state it; stacking hedges doesn't make an uncertain claim more honest, just softer.
• tells/copula-avoidance — "serves as," "stands as," "functions as," "represents a" substituted for plain "is/are." Named in Wikipedia's essay as a specific LLM tic; the dressed-up form adds syllables without adding meaning.
• tells/em-dash — every em dash, at error severity, not a rate check. A human reaching for a dash in running prose almost always reaches for a hyphen or a comma, so the character itself is the tell regardless of how often it appears. Use a comma, a period, or parentheses.
• tells/rule-of-three-density — a document-level check on "X, Y, and Z" triplet cascades. A rhythm device used once is fine; used as a reflex on every list, it reads templated.
• tells/title-case-header — headers capitalizing every word instead of sentence case. A formatting-convention tell, not a content one, but a strong signal on its own when combined with others.

When a violation looks wrong

If a genuinely load-bearing use of a "banned" word is being flagged (e.g. a technical document about a robust statistical estimator, where the word is a term of art, not filler), that's a signal the brief's register doesn't match this core's default general-audience contract — raise it to the user per hedgehog-copywriting-loop's rule on extending the contract, rather than working around the flag silently.