AI-Check Skill

Forensic analysis of text for AI-generation signals. Grounded in the published detection literature (Wu et al. 2025, Mitchell et al. 2023, Kujur 2025, AAAI 2025 shared task).

The output is a structured report, not a vague judgment. Every fired signal cites evidence.

The nine signal categories

Score each category 0–3:

• 0 = No signal detected (human-consistent)
• 1 = Weak signal (possible AI, could be human)
• 2 = Moderate signal (likely AI pattern)
• 3 = Strong signal (near-certain AI pattern)

Severity-to-score mapping (use for every category):

Double-counting policy: a single phrase can fire at most two distinct signals when the phrase is genuinely diagnostic for both. Example: "it is important to note that" is both Signal A (banned vocabulary) and Signal C (institutional hedge). Log it under both, but the same phrase cannot count as two separate weak instances inside the same category.

Total score cap: 9 categories × 3 = 27 maximum.

Signal A: Perplexity (word predictability)

Look for vocabulary that is maximally safe and expected — words that are technically correct but never the most precise or interesting choice a knowledgeable human would make.

Flags:

• Generic verbs where domain-specific ones belong ("address" instead of "untangle", "implement" instead of "wire up")
• Adjectives that describe without adding information ("significant improvements", "notable progress", "key challenges")
• Hedged assertions that swap specificity for safety ("can often lead to", "may result in", "tends to")
• Any of the canonical AI vocabulary list: delve, leverage (verb), utilize, robust, comprehensive, streamline, foster, facilitate, pivotal, nuanced, notable, notably, enduring, garner, it is worth noting, it is important to note, multifaceted, in the realm of, the landscape of, a myriad of, a plethora of

Cite the exact word or phrase that fired.

Signal B: Burstiness deficit (sentence uniformity)

Measure the variation in sentence length across the text.

Flags:

• Three or more consecutive sentences within 5 words of the same length
• No sentence shorter than 8 words in any 150-word block
• Metronomic rhythm — reading the passage aloud produces a steady pulse rather than natural variation
• No fragments used for emphasis

Report: list the sentence lengths in sequence (e.g. "14, 16, 13, 15, 17 — five consecutive sentences within 4 words of each other").

Signal C: Hedge density

Count the softening and epistemic hedge words.

Flags:

• "often", "generally", "typically", "in many cases", "it can be argued" appearing where direct assertion is warranted
• "it is important to note that", "it is worth mentioning", "one might consider"
• Diplomatic framing of obvious tradeoffs: "while X has benefits, it also presents challenges"
• Uncertainty expressed as institutional hedging rather than personal ("results may vary") vs human ("I'm not sure this holds when...")

Report: quote each hedge and note whether it was warranted by genuine uncertainty or reflexive softening.

Signal D: Structural tells

Look for document architecture patterns AI imposes regardless of content.

Flags:

• Bullet list where prose would serve better
• Topic sentence + evidence + restatement of topic sentence (humans skip the restatement)
• "In conclusion / To summarize / In summary" openers on closing paragraphs
• "In this [post/article/section] I will..." openers
• Numbered steps for content that isn't genuinely sequential
• Three-part structure imposed on every paragraph (intro, body, conclusion at micro-scale)
• Tricolon parallel structure: three examples or beats with identical grammatical shape e.g. "You X. Y. Does Z? You X. Y. Does Z? You X. Y. Does Z?" — perfectly symmetrical triplets in prose are AI-constructed. Real writers use two examples or vary the shape. Severity: strong.
• Perfect paragraph-per-idea arc: every paragraph does exactly one narrative job and advances the arc cleanly (setup → tension → lesson → evidence → reflection). Real personal writing has a paragraph that meanders, does two jobs, or doesn't fully resolve. A piece where every paragraph lands cleanly is architecturally perfect in a way human writing isn't. Severity: moderate in isolation, strong combined with other signals.
• Three-act Slack/update structure: for informal async messages, accomplishment → caveat → next steps maps directly to intro/body/conclusion. Real updates loop back, add a mid-message second thought, or end with something that doesn't fit the structure.
• Strawman pivot: "The case for X isn't about Y, it's about Z" / "It's not about X, it's about Y." Leading with what something is NOT before saying what it IS. Real writers lead with the actual point. Severity: moderate.

Signal E: Specificity deficit

Measure whether claims are grounded in concrete detail.

Flags:

• Abstract claim with no number, name, time reference, or example: "Many organizations have adopted..."
• Passive constructions obscuring the actor: "it has been found that", "research suggests"
• Universalist framing: "teams often find", "developers frequently encounter" (applicable to everyone, specific to no one)
• Named examples that are suspiciously generic or perfectly illustrative (AI picks canonical examples: "Netflix", "Amazon", "Stripe" without context)

Report: quote each unanchored claim.

Signal F: Transition word fingerprint

Catalog the connective tissue between sentences and paragraphs.

Flags (strong AI signals):

• "Furthermore," as paragraph opener
• "Moreover," as paragraph opener
• "Additionally," as paragraph opener
• "It is clear that"
• "This highlights / underscores / demonstrates the importance of"
• "As previously mentioned"
• "In addition to the above"
• "It goes without saying"
• "Needless to say"

Flags (moderate signals):

• "However," used more than once per 200 words
• "Therefore," used as a mechanical logical connector rather than earned conclusion
• "Turns out" / "it turns out that" as a pivot or reveal. AI uses this to create the illusion of a discovery narrative. "Turns out the config had a lower timeout" → "The config had a lower timeout." Quote each instance. Severity: moderate.
• Tutorial-voice transitions: "The standard fix is...", "The common approach is...", "Simple enough on paper" — these frame what follows as received wisdom, not personal experience. Strong signal in technical writing.
• Announcement-colon patterns: "The rule I use:", "The key insight:", "The approach here:", "The other thing I'd say:" — announcing before revealing. Severity: moderate. Also fires without a colon: "What I didn't expect was...", "What surprised me was...", "The thing I realized was..." — these are announcement sentences even without the colon. The colon isn't the tell; the announcement structure is.
• Pattern announcement: stating that a pattern exists before describing it. "The pattern is almost always the same" followed by the pattern. Real writers just describe the pattern.

Signal G: Punctuation fingerprint

Count the three AI punctuation tells:

Em dashes: Count total em dashes. More than 1 per 300 words is a signal. Specific sub-patterns:

• Double em dash wrapping (— like this —) is a near-certain AI pattern
• Em dash as pivot ("not mid-sprint — and the on-call rotation") — list-joiner em dash connecting two items within a sentence
• Em dash as dramatic aside ("X — which is worth noting — Y") Report exact count, location, and which sub-pattern.

Semicolons: Any semicolon linking two independent clauses in non-academic prose is a flag. Report exact count. Exception: comma-containing lists ("Austin, TX; Denver, CO").

Mid-sentence colons: A colon preceded by an incomplete clause ("The problem: nobody tests this" / "The answer: start earlier") is an AI structural pattern. Report each instance.

Signal H: Voice and register

Look for absence of human traces.

Flags:

• No first-person perspective anywhere in a piece where first-person would be natural
• No second-person direct address in instructional or opinionated content
• Consistent "polished neutral tone" — no personality variance, no roughness, no informality spikes
• No rhetorical questions used as transitions
• No self-correction or mid-thought qualification ("actually, that's not quite right")
• Opening sentence is a thesis, definition, or contextual framing rather than mid-thought or scene

Register collapse (Slack / informal writing): The most commonly missed signal in casual-register text. AI writes Slack messages that read like polished status reports with informal markers sprinkled in. Look for:

• Complete, well-formed sentences throughout — real Slack has fragments
• Topic-per-paragraph structure even in a short message
• Formal vocabulary underneath casual markers (~60% and lmk but the sentences themselves are well-constructed prose)
• No self-corrections mid-message ("oh also. just realized...")
• Three-act arc (accomplishment / caveat / next steps) intact beneath the informality
• Numbers written as words ("three incidents") rather than numerals with approximations ("~3 incidents", "<10min") Severity: strong when informal markers are present but prose structure is polished.

Templated closers in email/professional writing: "Happy to jump on a call if that's easier." "Let me know if you have any questions." "Feel free to reach out." These are the written equivalent of a throat-clearing opener. Real engineers end emails after the last substantive point, or with a specific ask, or with "lmk." Severity: weak in isolation, moderate when combined with other signals.

Signal I: Rhetorical scaffolding

Sentence and paragraph-level construction patterns that AI learned from polished writing and applies too consistently. These are the hardest signals to catch — they feel like good writing. Grammarly and live detectors flag these even when punctuation and vocabulary are clean.

Local coherence over-smooth (severity: moderate-strong, corpus-dependent) A pattern related to findings in recent research (DivEye, arXiv 2509.18880, TMLR 2026): every sentence connects too perfectly to the next, zero friction, zero cognitive-load artifacts. AI text often has no sentences that slightly misfire, no thoughts that shift direction mid-clause, no paragraph that doesn't close cleanly. Evidence: read each paragraph and check whether any sentence could be removed and the paragraph would still read perfectly. In human writing, removing a sentence often creates a noticeable gap. In AI writing, the paragraph usually flows better without it. Symptoms:

• Every paragraph opens with a claim and closes with a confirmation of that claim
• No sentence has a vocabulary mismatch with surrounding sentences
• No abrupt topic shift within a paragraph
• No sentence that slightly misfires before correcting

Calibration caveat (important). SHAP-based explainability analysis (arXiv 2603.23146) found that AI-text detectors rely on dataset-specific stylistic cues, not stable machine-authorship signals. Treat over-smoothness as a corpus-conditional indicator, not a universal authorship invariant. If the text is from a register that genuinely rewards tight coherence (academic abstracts, legal briefs, polished marketing copy), down-weight this signal.

Formula personal essay opener (severity: moderate) "The failure I think about most often happened in 2019." "The moment I remember most clearly was..." "The decision I regret most is..." Pattern: "The [noun] I [remember/think about/regret] most [adverb]" — AI's deliberate- introspection construction for opening personal essays. Real writers start with the incident, not with a ranked claim about their memory of it.

Asyndeton tricolon building in complexity (severity: moderate) Three items without conjunctions, each longer and more emotionally heavy than the last: "Two hours of degraded service, six engineers figuring out what I'd done wrong, a postmortem where I had to explain my reasoning to people who had been paged at home." AI constructs these to manufacture escalating emotional weight. Report the three items and their increasing length.

Intensifier/diminisher opposition (severity: moderate) "X [action] obsessively and Y [action] barely at all" — a balanced contrast using an amplifier against a diminisher. Same family as chiasmus but at the adverb level. Other forms: "X constantly / Y once", "X carefully / Y hardly". Quote the opposition.

Mini-aphorism paragraph closer (severity: moderate) A 4–7 word fragment or short sentence used to close a paragraph with a punchy lesson: "That's the part that stuck." "That's what changed." "That's the whole thing." "That's the real cost." AI appends these to tell the reader what conclusion to draw. Distinct from a sentence- length aphoristic closer — this fires even on very short fragments.

Landing phrase: "is the actual/real work" (severity: moderate) "Getting close enough to understand a failure is the actual work." "Deploying is the easy part. Debugging production is the actual work." AI's formulaic landing phrase for delivering conclusions. Quote it.

Parallel subject mirror (severity: weak-moderate) Two consecutive sentences opening with mirrored noun phrases that reflect each other: "The failure itself is just the event. Understanding it is separate." "The code is one thing. Maintaining it is another." AI constructs these as closing pairs. Report the mirrored subjects.

Participial reframe pivot (severity: moderate) Presenting a list of facts, then using a participial opener to reframe them as something more: "Laid out in a petition, the same facts read like a deliberate strategy." "Arranged that way, it sounds more planned than it was." "Seen this way, the whole arc reads differently." AI uses this pivot to manufacture the appearance of insight. The observation should be made directly without the reframing device. Quote the participial opener.

Thesis-first opener / "X is the easy/hard part" (severity: moderate) Starting a personal piece with the frame before the experience: "Gathering evidence for an EB1A petition is the easy part." "The writing is harder than the research." "X has become increasingly important." AI leads with the thesis because it's been trained on essays. Real writers start in the middle of the experience. Quote the opener.

Within-sentence anaphoric parallel list (severity: moderate-strong) Four parallel items with the same question-word structure inside a single sentence: "what existed before, what problem it solved, why the problem mattered, what changed after" Grammarly and other detectors score this identically to consecutive-sentence anaphora. The fix is varying the noun forms: "context, the actual problem, what changed" — not four parallel "what/why" question-clause starters. Quote the full list.

Composed self-aware parenthetical (severity: moderate) A parenthetical clause where the writer meta-comments on their own interpretation: "which I choose to read as progress" "which I take as a sign of X" "which I'm choosing to interpret as Y" These feel reflective but read as placed. Real reflection names the concrete behavior and stops; it doesn't append the writer's chosen interpretation of that behavior. Quote the parenthetical.

Parallel reason chains (severity: moderate) Three consecutive sentences with the same "subject + because/when + reason" structure, even when the subjects vary: "I filed patents because X. localaik started because Y. I gave talks when Z." The parallel clause shape is detectable even across different subjects. Vary the clause structure: one "because", one bare assertion, one gerund or fragment. Report how many parallel reason-chains fire in sequence.

"More X than Y" comparative framing (severity: moderate) All forms: "feels more like X than Y", "more specific than vague", "faster than". AI describes things by framing against an opposite. Humans describe directly. Quote the comparative.

"Not just X" / "not X, it's Y" / "not X but Y" diminishment (severity: moderate) Naming what something isn't before saying what it is. "It's not self-reported, it's merit-based." "Reasoning, not just behavior." All three forms are the same pattern. Quote the diminishment.

Setup sentences without colons (severity: moderate) Announcement sentences of the form "What [verb phrase] was [the revelation]" — the colon is not the tell, the announcement structure is. All forms fire:

• "What I didn't expect was..."
• "What surprised me was..."
• "The thing I realized was..."
• "What it didn't have was..."
• "What ended up working was..."
• "What changed everything was..."
• "What finally clicked was..."
• "What made the difference was..." Any sentence of the form "What [verb phrase] was [the revelation]" is an announcement sentence regardless of whether a colon follows. Quote the setup sentence.

Aphoristic / chiasmus closer (severity: moderate-strong) Two sub-patterns:

1. A closing sentence quotable as standalone: "The boilerplate is cheaper than the confusion." "The work doesn't sell itself."
2. Chiasmus — reversed parallel that sounds like insight: "Being specific about being wrong is more useful than being vague about being right." — "specific/wrong" mirrored against "vague/right." Real insight is asymmetric; AI constructs symmetric reversals. Quote and identify which sub-pattern.

Anaphora — same sentence-starter 2–3× consecutively (severity: moderate) "I still read slowly. I still lose the thread." "Why this structure. Why the error handling. Why the cache TTL." AI uses repeated openers for emphasis. Humans collapse them or vary the structure. Report the repeated opener and how many times it fires.

"Turns out" / "it turns out that" as reveal pivot (severity: moderate) AI's dramatic reveal device: "Turns out the config had a different timeout." Direct statement: "The config had a different timeout." The "turns out" adds nothing except the illusion of a discovery narrative. Quote each instance.

"Either X or Y" / "between X and Y" binary framing (severity: moderate) Clean binary choices presented as the only options. Real situations are a spectrum. "Teams face a choice between mocking (fast, but drifts) or live endpoints (accurate, but expensive)" — also fires the balanced parenthetical pattern below.

Balanced parenthetical pairs (severity: moderate) "(X, but Y) or (A, but B)" — two symmetric trade-offs in one sentence. Real trade-offs are asymmetric. The symmetry signals AI construction. Quote the parallel parentheticals.

Inverted burstiness (severity: weak) Three or more consecutive very short sentences (under 7 words each) without a longer counterweight. "The code was fine. The logic held. Nothing left a trace." Reads choppy in isolation. Distinct from Signal B which flags uniform medium-length sentences.

Mixed-authorship overlay (estimate how much AI editing)

In addition to scoring the 9 signals, estimate the AI-edited fraction — what portion of the text appears AI-written or AI-edited. This is a separate dimension from the verdict and addresses the real-world common case where humans edit AI drafts (or AI polishes human drafts). Framing borrowed from EditLens (arXiv 2510.03154), which regresses edit amount rather than predicting binary authorship.

Look for these distribution clues:

• Uniform AI signature across the whole text suggests pure AI generation
• Specific paragraphs polished, others rough suggests selective AI editing
• AI vocabulary clusters in transitions and conclusions, body is concrete suggests AI scaffold + human substance
• Opening / closing paragraph reads as polished, middle is fragmented or rough suggests AI editing of the bookends only
• Voice changes mid-text (formal → casual or vice versa) suggests mixed sources
• One paragraph has all the banned vocabulary and the others have none suggests a single AI-written section dropped into otherwise human text

Estimate the fraction in one of these buckets:

• Pure human (~0%)
• Lightly AI-assisted (~10-30%) — light polish, single section, or vocabulary substitution
• Mixed authorship (~30-60%) — substantial AI-written portions woven through
• Heavily AI-edited (~60-90%) — AI draft with human edits, or human draft with substantial AI rewriting
• Pure AI (~100%)

Report this as a separate line in the output format below.

Output format

Always output in this exact structure:

AI-CHECK REPORT
===============

VERDICT: [Human | Likely Human | Uncertain | Likely AI | AI]
CONFIDENCE: [Low | Medium | High]
OVERALL SCORE: [0–27] / 27
AI-EDITED FRACTION: [Pure human | Lightly AI-assisted | Mixed authorship | Heavily AI-edited | Pure AI]

SIGNAL BREAKDOWN
----------------
A. Perplexity            [0-3]  [one-line summary]
B. Burstiness            [0-3]  [one-line summary]
C. Hedge density         [0-3]  [one-line summary]
D. Structural tells      [0-3]  [one-line summary]
E. Specificity           [0-3]  [one-line summary]
F. Transitions           [0-3]  [one-line summary]
G. Punctuation           [0-3]  [one-line summary]
H. Voice / register      [0-3]  [one-line summary]
I. Rhetorical scaffolding [0-3]  [one-line summary]

EVIDENCE LOG
------------
[For every signal score > 0, list each specific instance with a short quote or description.
Format: SIGNAL-[LETTER] | "[exact quote or pattern description]" | severity: weak/moderate/strong]

WHAT GAVE IT AWAY
-----------------
[2–4 sentences identifying the strongest signals in plain language. Be specific about
which phrases, patterns, or absences were most diagnostic. This section is written
for a human who wants to understand the tell, not just see a score.]

RECOMMENDED FIXES
-----------------
[Only present if score > 6. Concrete rewrites or changes for the top 3 signals.]

Scoring thresholds

Calibration notes

• Short texts (<100 words) have fewer signals available; note this and adjust confidence to Medium max
• Technical writing with domain jargon can suppress Signal A even in AI text — don't penalize accurate domain vocabulary
• Academic or legal writing legitimately uses hedges and semicolons — adjust Signal C and G accordingly
• ESL writing can mimic some AI patterns (uniform sentence length, hedge-heavy); note if this is plausible
• A text can score AI on structure/transitions but human on voice — report both honestly
• Signal I (rhetorical scaffolding) fires on patterns that feel like good writing — do not discount them because the writing quality is high. These are the hardest tells precisely because AI learned them from skilled human writers. A "more like X than Y" comparative in an otherwise clean piece is still a signal.
• Register collapse (Signal H) requires cross-checking: informal markers alone do not make a Slack message human. Look at the sentence structure underneath the lmk and ~60%. If the underlying prose is polished and well-formed, the informal markers are surface noise.
• Aphoristic closers (Signal I) are context-dependent — a single well-turned closing line in a long personal essay is less diagnostic than the same pattern in a 200-word post where it's the only memorable sentence. Weight accordingly.

Known detection ceilings (cap confidence accordingly)

• Base-model output is a known ceiling. arXiv 2605.19516 ("Base Models Look Human") and corroborating Pangram analysis show that raw, non-instruction-tuned base-model output reads as human to current SOTA detectors. What modern detectors actually fire on is RLHF / instruction-tuning artifacts (polite hedging, structured enumeration, perfect coherence, "helpful assistant" register), not "AI-ness" per se. If the text plausibly came from a base model or a minimally-fine-tuned paraphraser (HIP-style attack), cap confidence at Medium even when surface signals look clean.
• Claude blind spot in zero-shot detectors. The DetectRL benchmark (arXiv 2410.23746) documents that Binoculars achieves only ~55% AUROC on Claude-generated text vs ~88% on GPT-3.5. If the source model is plausibly Claude, treat low scores with extra caution.
• Iteratively-paraphrased text is a ceiling. PADBen (arXiv 2511.00416) shows detectors >90% on direct AI text fail catastrophically on text that has been iteratively paraphrased through one or more LLMs. If the user mentions the text was paraphrased or rewritten, down-weight all signals.
• Stylistic cues are corpus-conditional. SHAP-based explainability analysis (arXiv 2603.23146) shows that surface stylistic features detectors rely on are dataset-specific, not stable authorship signals. This applies most strongly to Signal I (rhetorical scaffolding). Do not over-anchor on any single signal; require corroboration across categories.
• Multilingual text needs language-matched calibration. AI detectors badly misclassify non-English text — they wrongly flag lightly-polished human Arabic as AI, with one commercial detector dropping from 92% to 12% accuracy (arXiv 2511.16690). Refuse High confidence on non-English text unless calibration is known.

Reference detector landscape (for context)

If the user asks "what would tool X say?", these are the current characteristics:

• GPTZero (2025) uses RL adversarial self-training plus a learned classifier ensemble, not just perplexity + burstiness. Produces a 4-class output (human / slight / moderate / full AI-assist). Older "GPTZero relies on perplexity + burstiness" framing is stale.
• Binoculars is a strong zero-shot baseline but has the Claude blind spot above.
• Pangram 3.0 claims 99.98% accuracy with 1-in-10,000 FPR and 97% on humanized text per vendor benchmarks (independent replication pending).
• EditLens estimates AI-edit fraction rather than binary authorship (94.7 F1 binary, 90.4 F1 ternary).
• Ghostbuster is the canonical black-box (no token probs needed) detector — 99 F1 in-domain, degrades out-of-domain.
• DependencyAI uses syntactic dependency n-grams + LightGBM, cross-lingual without LLM access.

Humanize Text Skill

Transforms AI-generated or flat text into output that mirrors the statistical and stylistic fingerprint of human writing. Grounded in the published detection literature; sources live in references/research.md (background only, not needed during a rewrite).

Hard rules (read first, enforce last)

These seven fail more often than everything else combined, because the model that wrote the draft is the model checking it. You systematically overproduce these patterns; your draft contains em dashes even when you don't remember writing them. Treat "my draft is probably clean" as false by default.

1. Em dashes: at most one per 300 words of output. Under 300 words, zero.
2. Semicolons: none, unless a list item itself contains commas or the register is explicitly formal/academic (Lever 8).
3. Straight quotes and apostrophes only. Never curly. Sole exception: publishing contexts where typographic quotes are house style (Lever 8).
4. Banned vocabulary (full list at end of this skill). Highest-frequency offenders: delve, leverage, utilize, robust, comprehensive, streamline, furthermore, moreover, "it is important to note".
5. No negation framing: "not just X", "not X, it's Y", "it's not about X, it's about Y", "more X than Y". Say what the thing IS. Poetic forms count: "isn't proof you failed, it's proof you showed up" is the same banned pivot wearing a nicer coat. (False binaries like "either X or Y" are handled by the Signal I checklist's either/or item.)
6. Output shape: the rewritten text only. No preamble ("Here's the humanized version:"), no trailing changelog ("Main moves:", "What I changed:"). The ONLY permitted additions are the two meta-notes mandated by protocol steps 2 and 5.6, appended after the rewrite. If the user wants the diff explained, they'll ask.
7. Sentence-length spread: in any output over ~80 words, the longest sentence must beat the shortest by 20+ words, and fewer than half the sentences may sit in the 10-to-20-word band. Your uncorrected rhythm clusters at 10-20 words with ~6 words of deviation; that uniformity is a measured tell even when every other rule passes. Verify from the written count list (step 5), never by feel.

These apply in EVERY register, including creative, lyrical, and narrative prose. An em dash in a poem is still an em dash to a detector, and creative registers are where the "this one is doing literary work" rationalization kicks in hardest.

Enforcement is positional: stated here at the top, checked at the END against the finished draft by re-reading the actual draft text and counting occurrences (protocol steps 4-5). Never mark them clean from memory. If long context forces you to drop every other rule in this skill, keep these seven.

Mental model: what detectors actually measure

Nine signals: eight stylometric plus the RLHF fingerprint. Your output must move in the human direction on ALL of them.

The levers below are the write-side counterparts of the signals ai-check grades (A–I): 1→A, 2→B, 3→C, 4→D, 5→E, 6→H, 7→F, 8→G, 9→I (RLHF subset). The full rhetorical-scaffolding catalog for Signal I is enforced by the audit pass (step 5.5), not by any single lever.

Nine humanization levers, apply all of them

Lever 1: Perplexity injection (word-level)

Replace predictable vocabulary with words a real person would choose given this context:

• Swap generic verbs for specific ones: "address" → "untangle", "utilize" → "lean on", "implement" → "wire up"
• Let the subject matter suggest the vocabulary: a Go engineer says "flush the buffer", not "clear the temporary data storage"
• One or two genuinely surprising but accurate word choices per paragraph
• Avoid: "delve", "leverage", "robust", "streamline", "significant", "comprehensive", "notably", "it is worth noting", "in today's fast-paced world"

Watch for elegant variation (synonym cycling). LLMs cycle synonyms for the same referent: "The protagonist faces challenges. The main character must adapt. The central figure triumphs." Same person, three labels. Rule: pick the canonical noun per referent and use it consistently; vary with a pronoun, not a synonym. "the company / the firm / the organization" → "the company" + "it".

Lever 2: Burstiness injection (sentence-level)

Enforce sentence length variance. Target: standard deviation of sentence word count > 8. You can't compute stdev mentally, so enforce these two countable proxies instead; BOTH are required (hard rule 7):

• Range floor: longest sentence minus shortest ≥ 20 words. In practice: at least one fragment of 5 words or fewer AND at least one 25-plus-word sentence that earns its length.
• Mid-band cap: fewer than half the sentences in the 10-to-20-word band. Satisfying the range floor with one fragment and one long sentence while everything else sits at 12-16 words still reads uniform; the middle must spread too.

Supporting rules:

• Every 3–4 sentences, insert one sentence of 5 words or fewer. Just drop it. Like that.
• Never more than 3 consecutive sentences within 5 words of each other in length.
• Burstiness fails in both directions: a run of shorts without a longer counterweight reads choppy, not punchy.
• Mid-paragraph uniformity trap: openers and closers vary, but the middle 3–4 sentences collapse into the same band. Break one.

Lever 3: Hedge surgery

Audit every softening word:

• Delete: "it is important to note that", "it is worth mentioning that", "generally speaking", "in many cases", "it can be argued", "often", "typically" (unless genuinely needed for accuracy)
• Replace with direct assertion: "This matters because X" not "It is important to consider that X may be relevant"
• Real uncertainty gets human phrasing: "I'm not sure this holds for edge cases, but..." not "while results may vary"
• Don't soften real rules with "almost always" / "generally". If exceptions exist, name them: "This breaks down when X."
• No announcement-colon openers: "The rule I use:", "The key insight:" — state the rule directly. (Pattern announcement in general is a Signal I checklist item.)
• A sentence doing two logical jobs (comparison AND conclusion) usually reads cleaner as two.

Filler-phrase substitutions (the pattern generalizes: any multi-word wrapper around a one-word meaning gets the one word):

Lever 4: Structural flattening

Rhetorical scaffolding patterns (either/or binaries, chiasmus, tricolons, balanced parenthetical pairs, anaphora, "turns out" pivots, thesis-first openers incl. "X is the easy/hard part", mini-aphorism closers, parallel-subject mirrors) are catalogued ONCE in the Signal I checklist (step 5.5); negation pivots live in hard rule 5 and the step-4 diminishment scan. Apply the checklist at write time too. This table covers only what the checklist doesn't:

Lever 5: Specificity insertion

Every abstract claim needs a grounding anchor (a number, a name, a date, a concrete example). "Performance improved significantly" → "Latency dropped from 340ms to 80ms under the same load profile." If specifics aren't available, use plausible-specificity frames: "when you're running at X scale...", "in the cases I've seen...", "the one time this bit us..."

Lever 6: Voice and register

Human writing carries the writer's perspective:

• First-person where natural ("I find that...", "In my experience...")
• Occasional second-person direct address ("If you've ever debugged this...")
• Mild rhetorical questions as transitions: "So why does this matter?"
• Self-interruption mid-thought: "— actually, that's not quite right —", "more precisely:"
• Contractions in conversational registers: "don't", "it's", "you'll"

Lever 7: Discourse coherence (non-AI transitions)

Lever 8: Punctuation normalization

Em dashes (—). The most reliable single AI tell; AI uses them at 3–5× the human rate.

• Maximum one per 300 words; under 300 words, zero (hard rule 1)
• Only when a parenthetical genuinely interrupts rather than extends — not as a fancier comma
• Most uses replace cleanly with a period, a comma, or cutting the aside
• Never "X — like this — Y" (double em dash wrapping a clause); that pattern is almost exclusively AI
• X — item, item, item (introducing a list) → X. Item, item, item. or a colon after a complete sentence
• Three or more in one paragraph means structural problems; rewrite the paragraph

Semicolons (;). Real-world prose outside academic/legal writing almost never uses them.

• Treat every semicolon as a bug unless the register is explicitly formal/academic
• Replace with a period (usually), "and"/"but"/"so" (when the relationship matters), or restructure
• Exception: lists whose items contain commas ("San Francisco, CA; Austin, TX")

Mid-sentence colons (:). Fine at the end of a complete clause to introduce; mid-thought is an AI pattern.

• "The answer is: start earlier" → "Start earlier."
• "The problem: nobody tests this" → "Nobody tests this." or "Here's the problem: nobody tests this."
• One colon per paragraph maximum in non-list prose

Curly quotes. A near-certain single-character tell that survives rewriting.

• Find-replace before shipping: curly “ ” → straight ", curly ‘ ’ → straight ' — apostrophes included
• Exception: publishing contexts where typographic quotes are house style

Lever 9: Strip RLHF / instruction-tuning voice

Current detectors mostly fire on RLHF and instruction-tuning artifacts, not "AI-ness" per se ("Base Models Look Human"; details in references/research.md). What gets flagged is the "helpful assistant" voice. This lever is the single most valuable one. Strip:

Advanced techniques (optional, when stakes are high)

The nine levers are pure-rule; hybrid (rule + model-in-the-loop) approaches benchmark better. When stakes warrant the cost, layer these on (sources: references/research.md):

1. Detector-scored best-of-N. Generate 3–5 variants; score each against a real detector (GPTZero, Pangram, Binoculars), a banned-word count, or a perplexity probe; ship the lowest.
2. Iterative paraphrase pass. Run the output through a second LLM with "paraphrase this, keep the meaning." Diminishing returns past 2 passes; meaning drift accumulates — verify substance.
3. Writer-profile distillation. When the user supplies writing samples, distill style hypotheses first (this is protocol step 0; follow the procedure there). Beats raw few-shot.
4. Self-rewrite distance check. Ask a different LLM to "rewrite this in different words." Near-identical rewrite = text still at a local probability maximum = reads as AI.
5. Embedding-guided synonym swap. When tooling is available, prefer substitutions that explicitly lower detector scores over Lever 1's static word list.
6. Disfluency injection (casual register only). Light hesitations, mid-thought restarts, "wait actually" corrections. Off by default for formal writing; disfluencies in a board memo are their own tell.

Dead ends, don't bother: homoglyph injection (defeated by Unicode normalization, and a clear tampering signal), single cross-model rewrite (doesn't defeat trained detectors alone), watermark stripping (separate problem space).

Rewrite protocol

When given text to humanize:

0. (Optional) Writer-profile distillation. If the user provided prior writing samples, extract style hypotheses across six dimensions before touching the new text:

   1. Sentence length pattern. Variance? Signature fragments?
   2. Word choice level. Casual or academic? "stuff"/"thing" or "elements"/"components"?
   3. Paragraph openers. Straight in? Context first? A question? A scene?
   4. Punctuation habits. Em dashes? Parentheticals? Ellipses? Fragments?
   5. Recurring phrases / verbal tics. Repeated phrases? Fillers ("honestly", "basically", "look,")?
   6. Transition style. Explicit connectors, or next thought with no bridge?

   Distill 5–10 specific hypotheses ("never opens with a thesis", "fragments in conclusions", "sentence variance roughly 6–28 words").

   Critical rule when matching voice: don't just remove AI patterns — replace them with patterns from the sample. If the sample is casual, don't upgrade the vocabulary. The skill's default bias toward terse, direct prose yields to the sample's register when they conflict. Then apply the levers in service of those hypotheses.

1. Read the full input first. Identify topic domain, audience, register, length target.

2. Inventory the AI tells. Flag hedge count, list/bullet count, sentence-length uniformity, examples present/absent, transition inventory, RLHF voice markers.

   Also count specific anchors (numbers, named entities, dates, time references, concrete examples). If the count is zero AND no voice sample was provided in step 0, still humanize — use Lever 5's plausible-specificity frames, never invented facts — then append this note AFTER the humanized text, blank-line separated, as plain text (no > marker):

   [Note: the input had no factual anchors (no numbers, names, dates, or specific examples). The rewrite is cleaner but learned classifiers (GPTZero, Grammarly) may still flag it on the specificity signal alone (Signal E in ai-check). To close that gap, give me the actual specifics (product names, metrics, dates, named tools) or a sample of your writing to match.]

   Do not stop and ask before rewriting. By the time the user says "proceed anyway", this skill sits deep in the conversation history and the second-pass rewrite reliably leaks tells back in. Rewrite now, flag the gap after.

3. Rewrite in a single pass applying all nine levers. No "light editing"; the statistical fingerprint requires structural change.

   This matters most when the input is text you wrote earlier in this conversation. Rewriting your own recent output anchors you to its phrasing, and the rewrite silently degrades into word swaps that leave the original's em dashes, negation pivots, and rhythm intact. Treat your own prior output as foreign text: extract what it says, re-derive the prose from the content. If your edit log would read as a list of substitutions, you light-edited. Start over.

4. Pre-output gate. A literal scan, not a recollection: re-read the draft top to bottom and for each item write the count and quote every hit before fixing it. Write zeros explicitly ("em dashes: 0"). A gate entry without an explicit count is a gate you did not run — an unenumerated "looks clean" always passes, and this is where humanization fails silently in practice.

   • Em dashes. Scan for "—", write the count. More than (word_count / 300)? Cut or replace with periods.
   • Semicolons. Scan for ";", write the count. Replace with a period or "and"/"but"/"so" unless a comma-containing list.
   • Curly quotes/apostrophes. Scan for the curly characters “ ” ‘ ’, write the count. Replace with straight " and '.
   • Banned vocabulary. Scan against the master list (end of this skill), quote each hit. Eyeball first: delve, leverage (verb), utilize, robust, comprehensive, furthermore, moreover, "it is important to note".
   • Comparative framing. Scan for "more ... than" and "feels like ... not", quote each match. Describe the thing directly.
   • Diminishment. Scan for "not just", "not X, it's", "not X but", quote each match. State what it IS.

   After fixing hits, re-scan the sentences you rewrote: regenerated prose reintroduces the same tells at the same rate as the first draft.

5. Self-check. Step 4 cleared the mechanical tells; this covers the rest:

   • Sentence rhythm, from counts, not feel. Write out every sentence's word count in order ("9, 5, 22, 16, 7..."), then check the list against ALL four, fixing and recounting until every one passes (hard rule 7 covers the first two; conditions 1-2 apply only to outputs over ~80 words):
     1. Max minus min ≥ 20 (needs a ≤5-word fragment and a 25-plus-word sentence)
     2. Fewer than half the counts in the 10-to-20 band
     3. No three consecutive counts within 5 words of each other
     4. At least one count ≤ 6 per 150 words of output A mental read-through always sounds varied to the model that wrote it; the number list doesn't lie. Standard fixes: split one mid-length sentence into a fragment plus the remainder, and merge two mid-length neighbors into one long sentence that earns it.
   • Every paragraph has at least one specific anchor (number, name, example, time reference). On the step-2 zero-anchor path, Lever 5's plausible-specificity frames satisfy this item — never invent facts to pass it.
   • No bullet lists unless the user requested them
   • Voice consistent throughout (no third-person formal → first-person casual drift)
   • Every colon preceded by a complete sentence (exception: Slack/informal fragments)
   • Output shape (hard rule 6). No preamble, no trailing changelog. Only permitted additions: the meta-notes mandated by steps 2 and 5.6.
   • Rhetorical scaffolding: scan for the Signal I checklist patterns (step 5.5) — they survive the gate because they feel like good writing. Outputs >150 words get the full list in the audit; shorter outputs at minimum check aphorism closers, anaphora, "turns out" pivots, "What X was Y" setups, either/or binaries.

5.5. Audit pass. Run the Signal I checklist below on every output — it is the single source of truth for rhetorical-scaffolding patterns. For outputs >150 words, also run the rewrite-and-recheck loop: after the self-check passes, ask "What still makes this read as AI?", list 2–3 residual patterns, rewrite those sentences, re-run the gate and self-check. Empirically the first revision has 2–4 Signal I patterns left; one loop gets it to 0–1. Loop once only — past iteration 2 you over-edit into choppy, voiceless prose.

Flagged residuals must be removed, not justified. Not kept because "removing it would collapse the paragraph", "this register needs it", "it reads thin without it", "it would be choppy", or "it's a transition, not a closer". Those rationalizations are how Signal I patterns survive — they feel necessary because they're constructed to feel necessary. If removing a flagged sentence makes a paragraph too thin, the paragraph IS too thin: collapse or merge it. An honest 80-word output beats a padded 200-word output that reads as AI. Red flag: if your audit says "borderline but I'm keeping it because...", you just lost the loop. Cut it.

Signal I checklist (every audit, every paragraph). A general "what reads as AI?" prompt misses things; scan for each pattern and fix every hit:

• Mini-aphorism closer. Paragraph ends in a 4-to-10-word punchy "lesson" ("That's the part that stuck.", "Slide decks don't.")? Cut it, or fold it into the previous sentence.
• Thesis-first paragraph opener. Frame before experience ("The rollout was the hard part.", "The real question is Y.")? Start with the concrete thing; let the thesis emerge.
• Parallel-subject mirror. Consecutive sentences with mirrored noun-phrase openers ("The code is one thing. Maintaining it is another.")? Vary one subject.
• Aphoristic closing sentence on the whole piece. Last sentence reads as quotable standalone wisdom? End on a specific detail, an unresolved question, or a concrete next action.
• Pattern announcement. Naming a pattern before describing it ("The pattern is X.")? Just describe it.
• "Turns out" pivot. Reveal-narrative framing? State the discovery directly.
• Setup sentences. "What I didn't expect was X."? Lead with X directly.
• Within-sentence anaphoric parallel list. Four parallel question-word items ("what X, what Y, why Z, what W")? Use varied noun forms instead.
• Composed self-aware parenthetical. Meta-commentary on your own state ("which I choose to read as X")? End on the concrete behavior.
• Parallel reason chains. Three consecutive "subject + because/when + reason" sentences? Vary the clause structure: one "because", one bare assertion, one fragment.
• Anaphora. Same opening word on 2+ consecutive sentences? Collapse or vary the opener.
• Either/or binary. Clean two-option framing of a spectrum? Name the actual situation.
• Balanced parenthetical pairs. Symmetric trade-offs in one sentence ("(X, but Y) or (A, but B)")? Real trade-offs are asymmetric; break the symmetry.
• Tricolon. Three parallel beats with identical grammar, often conjunction-free and escalating in weight ("Two hours of X, six engineers doing Y, a postmortem where Z.")? Break the third item into its own sentence, join two with "and", or reduce to two.
• Chiasmus / balanced opposition. Reversed-parallel construction that sounds like insight ("being specific about being wrong / being vague about being right")? Make the comparison asymmetric.

3+ hits means the patterns compound — address all of them. Two mini-aphorisms might be tolerable; three in five paragraphs is a clear AI signature.

5.6. Output-length sanity check. If output is under 50% of input length, the input was mostly puffery that got correctly removed. Don't pad it back up — padding reintroduces the stripped patterns. Instead append, blank-line separated, as plain text (no > marker):

[Note: input was substantially puffery; humanized output is N% shorter. To make this longer without re-introducing AI patterns, add specific anchors: numbers, named entities, examples, or time references.]

This and step 2's note are the only commentary the skill ever outputs, always after the rewrite, clearly separated. The intent: make the gap visible instead of silently shipping thin output that fails on Signal E.

6. (Optional) Self-rewrite distance sanity check. When stakes are high, run Advanced 4.

7. (Optional) Detector-scored best-of-N. When stakes are high, run Advanced 1.

8. Output the rewritten text only. No preamble ("Here is the humanized version:"), no trailing changelog ("Main moves:", "What I changed:"). The only permitted additions are the meta-notes mandated by steps 2 and 5.6. This holds in chat interfaces too, where narrating edits feels helpful: it isn't the deliverable, and a changelog of word swaps is evidence you light-edited (step 3). If the user wants a side-by-side, they'll ask.

Generating human text from scratch

When writing new content (not rewriting):

• Apply all nine levers from the first sentence
• Start mid-thought when it fits: "The tricky part about X isn't what most people think."
• No "In this [article/post/section], I will..." opener
• End without a summary paragraph unless the piece is long enough to genuinely need a re-anchor
• Calibrate voice to the domain: an engineer's Slack post sounds different from a board memo

Decoding-strategy note (when controlling generation): set temperature high (0.9–1.1), top-p loose (0.95–0.99), repetition penalty up (1.1–1.2). This widens the token distribution and breaks the local-maximum property perplexity detectors rely on (RAID benchmark; see references/research.md).

Domain-specific calibration

Technical (engineering, code, systems)

• Domain-native vocabulary: "the hot path", "this falls apart at scale", "the footgun here is..."
• Short sentences for definitive claims: "This is O(n²). Don't do it at scale."
• Tradeoffs direct, not diplomatic: "The downside is real: you lose..."
• Real tool names, version numbers, error messages when available

Narrative / blog / essay

• Open with a scene or specific moment, not a thesis
• Let the argument emerge from the evidence
• Sentence fragments deliberately for rhythm. Like this.
• One moment of genuine uncertainty or changed mind per 500 words

Creative / lyrical prose (fiction, poetic passages, mood pieces)

The register where this skill gets rationalized away. Every hard rule still applies; detectors don't grade on artistic merit. The traps:

• The "literary em dash" exemption. "Something wouldn't let you stay asleep — a feeling too new to have a name yet." Feels earned; still the single strongest tell. Use a period, a comma, or restructure.
• Poetic negation pivots. "It isn't proof you failed — it's proof you showed up." Say the positive thing: "It's proof you showed up."
• Anaphora as lyricism. "Not every door has been tried. Not every version of yourself." Reads as AI cadence, not poetry. Vary the second sentence.
• Mid-sentence colon reveals. "and you think: this is where it starts" — restructure or end the sentence before the colon.
• Balanced imagery pairs and escalating tricolons. Human lyrical prose is lopsided; AI lyrical prose is symmetric. Break the symmetry.

Human creative writing gets its texture from specificity and asymmetry (a named street, a wrong note, an image that doesn't resolve), not punctuation drama.

Professional / business

• Cut the throat-clearing opener; just the message
• State the ask in sentence 1 or 2
• Numbers and deadlines: "by Thursday EOD", "the three blockers are..."
• Short paragraphs (2–3 sentences max in email/memo)

Slack / async team updates

Register collapse is the primary tell: AI Slack reads like a polished status report. Real Slack has:

• Abbreviations and approximations: ~60%, <10min, fwiw, btw, lmk, tmrw
• Fragments throughout: "hard commits: billing gRPC + pprof thing", not full clauses
• Self-corrections mid-message: "oh also —", "wait, actually"
• Thoughts bleeding together, looping back, trailing off — not topic-per-paragraph
• Numerals with approximation markers: ~3-4 days, not "approximately three to four days"
• Single-word endings: "lmk" as its own line
• Lowercase throughout except proper nouns
• The structure must actually break: accomplishment → caveat → next steps is still AI even with fwiw sprinkled in. Add a fourth element that doesn't fit, loop back, or end with an unset-up question.

What this skill does NOT do

• Guarantee 0% AI scores on commercial detectors (no method does reliably)
• Add false information to increase specificity — plausible framing only
• Change the factual content of the input, only the expression
• Apply the same transformation to every domain — register matters

Reference: banned word/phrase list (compile-time errors in AI text)

Remove every instance before outputting:

Core AI vocabulary: delve, leverage (verb), utilize, robust, comprehensive, streamline, foster, facilitate, pivotal, nuanced, multifaceted, crucial (overused), enduring, garner, valuable, vibrant, tapestry (figurative), testament (figurative), interplay, intricate, intricacies, landscape (as abstract noun), showcase (verb), highlight (as standalone verb), underscore (as standalone verb), align with, actually (as filler), additionally (as opener)

Hedge / softener clusters: it is important to note, it is worth mentioning, notably, it's worth noting, in many cases, generally speaking, it can be argued

Filler / formula openers and closers: in today's fast-paced world, in conclusion, in summary, to summarize, it goes without saying, needless to say, at the end of the day, at its core, under the hood, the standard fix, the common approach, simple enough on paper

AI transition fingerprint: furthermore, moreover, it is clear that, this highlights, this underscores, as previously mentioned, turns out (as a pivot), it turns out that

Significance inflation: stands as a testament to, marks a pivotal moment in, indelible mark, evolving landscape, setting the stage for, deeply rooted in, plays a vital role, a key turning point, represents a shift in

Promotional / marketing register: nestled in the heart of, in the heart of, breathtaking, must-visit, stunning, boasts a rich heritage, renowned for, groundbreaking (figurative), vibrant (cultural copy)

Quantifier inflation: a myriad of, a plethora of, in the realm of, the landscape of (abstract)

Persuasive authority tropes: the real question is, what really matters, fundamentally, the deeper issue, the heart of the matter, in reality

Signposting / tutorial scaffolding: let's dive in, let's explore, let's break this down, here's what you need to know, now let's look at, without further ado

Knowledge-cutoff disclaimers: as of my training cutoff, up to my last training update, while specific details are limited based on available information, based on what I know up to

Sycophantic prefixes: great question, you're absolutely right, that's an excellent point, of course!, certainly!

Templated email / Slack closers: happy to jump on a call, let me know if you have any questions, feel free to reach out, i hope this helps, looking forward to connecting soon

Binary framing: whether X or Y (as a clean binary framing opener)