AI tells

A tell is an involuntary giveaway. Not a rule violation, a habit that leaks the writer's nature. You are hunting tells, not scoring "AI-ness": detectors are unreliable, humans score near chance on this task, and no single tell is proof. Density is the signal. Three tells in a paragraph is a rewrite. One em dash is nothing.

Distilled from Wikipedia:Signs of AI writing (read 2026-07-24), filtered to what applies to first-person prose. Its wikitext, DOI/ISBN, category, template, and edit-summary sections are Wikipedia-specific and deliberately dropped.

Run order

1. Grep pass — mechanical, catches the lexical tells.
2. Read pass — the structural tells no grep sees (§ Read pass).
3. Rewrite, then re-grep. Fixing one tell routinely plants another: killing repetition invites synonym-churn, killing "serves as" invites "functions as".

Done when: grep is clean or every surviving hit is a named deliberate keep, every Read-pass item has been walked one at a time, and the prose still sounds like its author rather than like nobody.

Grep pass

Two passes, because hard-wrapped prose splits multi-word phrases across newlines — in a file wrapped at ~70 characters, rather than is invisible to a line-based grep about half the time. Run pass B whenever the target is wrapped; skip it only when the file is one paragraph per line.

F=path/to/draft.md   # set to the file under audit

# A1. Model junk markers. Any hit is unconditional deletion.
grep -noE 'contentReference|oai_?citation|oaicite|turn[0-9]+(search|view|image)|attributableIndex|\[cite: |grok_(card|render)|ppl-ai-file-upload|attached_file|:::writing' "$F"

# A2. Lexical tells, as a frequency table. Read the counts, not the list.
grep -nioE '\b(serves? as|stands? as|functions? as|operates? as|boasts?|underscor(es?|ing)|showcas(es?|ing)|emphasiz(es?|ing)|highlight(s|ing)|foster(s|ing)|enhanc(es?|ing)|leverag(es?|ing)|pivotal|crucial|vital role|testament|tapestry|delve|intricate|meticulous|vibrant|robust|seamless(ly)?|enduring|realm of|landscape|navigat(e|ing|ion))\b' "$F" \
  | cut -d: -f2- | tr 'A-Z' 'a-z' | sort | uniq -c | sort -rn

# A3. Vague attribution. Collides with fact-grounding; treat every hit as a sourcing bug.
grep -niE '\b(experts? (say|argue|agree|note)|industry reports?|observers have|some critics|studies show|research shows|widely (held|regarded|considered))\b' "$F"

# A4. Typography. Count em dashes against the deliberate keeps below.
printf 'em dashes: %s\n' "$(grep -o '—' "$F" | wc -l | tr -d ' ')"
grep -nE $'[‘’“”]' "$F"   # curly quotes/apostrophes: match the file's own convention

# B. De-wrapped phrase pass. No line numbers, so it answers "is it in here",
#    then you locate it by searching the phrase.
tr '\n' ' ' < "$F" | grep -oiE \
  "not (only|just|merely|simply)[^.]{0,60}\bbut\b|\b(it|this|that)(.s not|( i)?sn.t) [^.]{0,40}[.,] +it.s|\bnot a [^.]{0,40}, not a\b|\brather than\b|despite (its|the|this)[^.]{0,80}(challenge|obstacle|hurdle)|faces (several|numerous|various)|\bnavigate the landscape\b|in today.s [^.]{0,20}world|it.s worth noting|in conclusion" \
  | sort | uniq -c | sort -rn

Vocabulary ages. The 2023-era set (delve, tapestry, testament, meticulous, intricate) now flags old-model output specifically. Post-mid-2025 the tell narrows to four survivors: emphasizing, enhance, highlighting, showcasing. Weight a hit on those four heaviest; they are what current models still overproduce.

Tells, ordered by how often they bite

1. Copulative dodge. is and are replaced by marketing verbs: serves as, stands as, functions as, boasts, features, offers, marks. Highest-frequency tell in the corpus. → Write is. It is not a weaker verb, it is the correct one.

2. Negative parallelism, three shapes, all tells: not only X but also Y · not X, it's Y / not X. It's Y · X rather than Y. The third is the one people never notice and models never stop writing. All three fake a misconception the reader never held. → State the positive claim alone. If the contrast is real, give the rejected option its own sentence.

3. Significance inflation. Mundane subject wrapped in importance: underscores, reflects a broader, pivotal/crucial/vital role, setting the stage, symbolizing, enduring legacy, testament to. → Cut the significance sentence entirely. If the thing matters, the specifics already said so.

4. Participial fake-analysis. Trailing -ing clause that asserts analysis without doing any: "…, highlighting the importance of trust", "…, fostering community", "…, contributing to growth". → Either explain the mechanism in its own sentence or delete the clause. It carries no information in either case.

5. Vague attribution. experts argue, industry reports, observers have noted, some critics, studies show, widely held. → Name the source, or make it a first-person claim from a seat you actually hold, or cut it. This one is also a fact-grounding failure, not only a style tell.

6. Formula conclusion. "Despite its [positives], X faces several challenges, including A, B, and C" closing on vague optimism. Also the headings that host it: "Challenges and Legacy", "Future Outlook". → End on the most specific thing you know, or on what you would do next.

7. Rule of three. Triplet adjectives and exactly-three-item lists, everywhere, regardless of how many things there are. Ungreppable, so it survives every mechanical pass. → Count the items honestly. Two is two, five is five. A triplet you'd have written anyway is fine; a triplet per section is a tic.

8. Synonym-churn (elegant variation). The same noun renamed on every mention to dodge repetition — "artistic constraints", "non-conformist artists", "their creativity" all pointing at one subject in one paragraph. → Repeat the word. Repetition reads human; churn reads generated.

9. Promotional register. vibrant, rich, profound, boasts, nestled, groundbreaking, renowned, diverse array, breathtaking. Travel-brochure tone in a first-person piece. → Say the concrete thing the adjective was standing in for.

10. Formatting habits. Title Case headings where the catalogue uses sentence case · bolding every key term on every appearance · bullet lists shaped **Header:** description · emoji as visual structure · tables for what is really prose · thematic breaks (---) above headings · em-dash pileups. → Match the neighbors in the same directory. The tell is divergence from the corpus, not the mark itself.

Deliberate keeps

Every corpus has marks that look like tells and are not. Get the keeps before the read pass — from the caller, from the skill that mandates the format, or from the neighbouring files. Name them up front so they stop being findings on every run.

Three recur in almost any first-person corpus:

• Em dashes are not banned. A mandated signoff line and a disclaimer cadence are voice, not tells. Judge em dashes by rate, never by presence: a couple per piece plus the signoff is voice, six in three paragraphs is a tell.
• Curly quotes are correct wherever the surrounding corpus uses them. They are a tell in wikitext, not in Markdown prose. Match the file.
• First-person "I" is a conversational-AI tell only where the format forbids it. Where the voice is mandated first-person, it is the voice.

Declared keeps for a specific corpus live in references/corpus-<name>.md. Read the one matching the file under audit, if it exists — it names that corpus's signoff, quote convention, and any format the caller mandates that would otherwise read as a tell. No reference file means ask the caller.

The GEO tension, resolved

A caller optimizing for AI search will prescribe question-shaped H2s with a self-contained, conclusion-first lead answer in the first 40-75 words. That skeleton is structurally the "outline-like, formulaic section" tell, and it is kept on purpose: AI search engines extract at the passage level.

Resolution: keep the skeleton, forbid the filler. Every lead answer must carry a number, a stake, a lived detail, or a counter-take. A lead answer that neutrally restates the heading is the tell arriving through the front door. Where a corpus reference declares such a skeleton, it is a keep; the filler inside it never is.

Not this skill's job

No AI-likelihood score, no percentage, no detector. Those are the things the source page documents as unreliable. Output is a located list of tells and their rewrites, or "clean".