humanize

name

humanize

description

Removes AI-writing tells from prose and prevents them when drafting: inflated significance, hedged comparisons, chat-register leakage, uniform rhythm, formulaic structure, machine residue. Invoke whenever task involves prose meant for human readers — writing or editing articles, blog posts, READMEs, documentation, release notes, announcements, reports, or reviewing text suspected of sounding AI-generated.

Humanize

Fix the writing, not the detection. AI-sounding prose is bad because it is vague, padded, and evasive — not because a detector might flag it. Every edit must improve the text for a human reader. Edits whose only purpose is dodging AI detectors are out of scope.

Two failure modes carry equal weight:

• Under-editing — tells survive, the text reads as generated slop.
• Over-editing — voice, specifics, and legitimate style get flattened into different slop. Gutting a human quirk is as bad as leaving a machine one.

Hard Constraints

• Never fabricate. Rewrites never invent facts, numbers, dates, names, quotes, anecdotes, or first-person experience. When the fix for vagueness is a concrete detail you don't have, get it from the source material, ask the author, or keep the sentence plain. A vivid lie is worse slop than a dull truth.
• Conserve substance. Every claim in the original survives the rewrite unless the author asks for cuts. Rewrite, don't summarize: a five-paragraph original yields a rewrite that covers all five paragraphs' content.
• Match the voice. Fit the document's intended register (formal, casual, technical). If a writing sample of the author exists, mirror its sentence rhythm, word level, and punctuation habits instead of installing your own.
• House style wins. When the surrounding document or project style uses a pattern deliberately, match it. The skill fixes unexamined defaults, not deliberate style choices.
• Secondhand text is immune. Never rewrite quotations, titles, proper names, or text that is being discussed rather than used.

The Mechanism Model

Tells are symptoms of a few generative mechanisms. Learn the mechanisms and you catch variants no list contains. Each family below gives the mechanism and its triggers; per-pattern before/after examples live in the reference.

1. Inflated importance

LLMs regress to the mean: rare specifics get replaced with generic significance. The subject becomes simultaneously less specific and more exaggerated — "inventor of the first train-coupling device" fades into "a revolutionary titan of industry".

• Significance puffery — "stands as a testament", "pivotal moment", "marking a shift", "evolving landscape", "indelible mark", "setting the stage for". Fix: state what the thing is or does; delete the significance claim or replace it with the specific fact it gestures at.
• Superficial -ing analysis — a participle clause tacked onto a sentence to add fake depth: ", highlighting...", ", ensuring...", ", reflecting...", ", underscoring...". The single strongest 2026 tell by corpus data ("ensuring" is 4.3× over-represented in AI text). Fix: cut the clause, or promote it to its own sentence with a named actor and a verifiable claim.
• Promotional tone — "vibrant", "nestled", "breathtaking", "rich heritage", "must-visit", "boasts". Fix: neutral description with concrete attributes.
• Symbolic gloss — narrating the meaning of a fact instead of trusting it: "represents", "symbolizes", "embodies", "speaks to" applied to mundane things. Fix: state the fact; let the reader interpret.
• Vague authority — "experts argue", "industry reports", "observers have noted", "widely regarded". Fix: name the source and what it actually said, or cut the claim.
• Notability name-dropping — lists of outlets or follower counts as proof of importance. Fix: one source, one specific statement from it.
• Generic upbeat conclusions — "the future looks bright", "exciting times ahead", formulaic "Challenges and Future Prospects" sections. Fix: end on the strongest specific point; concrete plans beat vibes.

2. Performed deliberation

The model performs considered-ness instead of committing to a claim. Rhetorical shapes stand in for actual thought.

• Hedging verbs as padding — "ensures", "supports", "reflects", "highlights" gluing an idea to an unearned benefit. Fix: say what the thing does, with evidence or not at all.
• Hedged comparison — "X rather than Y", "X, not Y" where the contrast adds nothing. The strongest multi-word tell in 2026 corpus data (2.5× over-represented). Fix: make the comparison (state why X) or drop Y entirely.
• Negative parallelism — "It's not just X, it's Y", "not only... but also", and clipped tailing negations ("no guessing", "no wasted motion"). Fix: assert the positive claim as a real clause.
• Contrast-frame pileup — "it is X, not Y" recurring through a text. One instance with a real contrast earns its place; the frame is a default sentence shape for every current model, so two or more in the same passage convict even when each is individually defensible. Fix: keep the strongest one, restate the rest as direct claims without the foil.
• Rule of three — ideas forced into triads for fake comprehensiveness. Fix: keep the items that carry weight, one, two, or four as reality dictates.
• Intensifiers without evidence — "significantly", "effectively", "seamlessly", "dramatically". Fix: back it with a number or delete it.
• The role formula — "plays a crucial/key/vital role in shaping..." — statistically the single most formulaic sentence shape LLMs produce. Fix: replace with the specific action: what does it do, to what, with what result.
• False ranges — "from X to Y" where X and Y aren't on a scale ("from the Big Bang to dark matter"). Fix: list the actual items or name the real dimension.
• Aphorism formulas — "X is the Y of Z", "the currency of", "not a tool but a mirror". Fix: state the concrete claim the aphorism gestures at.
• Fake-profound framing — "The real question is", "at its core", "what really matters". Fix: delete the frame; the sentence that follows usually stands alone.
• Hedged enumeration openers — "There are several ways to...", "Generally speaking,". Fix: give the answer first.
• Excessive hedging — stacked qualifiers: "could potentially possibly". Fix: one qualifier, or commit.

3. Leaked context

Text meant for the conversation, the task, or the model's own scaffolding bleeds into the artifact.

• Chat correspondence — "I hope this helps", "Certainly!", "Would you like me to...", "let me know". Fix: delete; start with the content.
• Sycophancy — "Great question!", "You're absolutely right". Fix: delete.
• Signposting — "Let's dive in", "Here's what you need to know", announcing instead of doing. Fix: do the thing.
• Engagement hooks — "The catch?", "Here's the kicker.", "Sound familiar?", fake-candid openers ("Honestly?", "Look,"), and runs of staccato drama fragments. Fix: deliver the point without the theatrical pause; one short sentence for emphasis is fine, a run of them is engineered.
• Reasoning scaffold — "Let me think through this", "Step 1:", "Breaking this down" left in final text. Fix: keep the conclusion, delete the scaffolding.
• Knowledge-cutoff residue and gap-filling — "as of my last update", "while details are scarce...", "maintains a low profile", "likely grew up...". Fix: say what is known with a source, say what isn't known plainly, or cut. Never dress a guess as fact.
• Diff-anchored writing — docs narrating the change instead of the current state: "was added to replace", "now uses", "has been updated to". Unless the document is version-scoped (changelog, migration guide), describe the thing as it is.
• Fragmented headers — a heading followed by a one-line warm-up restating it. Fix: cut the warm-up; start with content.

4. Uniform texture

Statistical generation produces even, interchangeable prose: same-length sentences, cycled synonyms, restated ideas, paragraphs that don't build on each other.

• Flat cadence — every sentence the same length and shape. Fix: vary deliberately. Short sentences land points. Longer ones carry qualifications and let a thought wind to where it's going. Read the paragraph aloud; if it thuds evenly, reshape it.
• Synonym cycling — "the protagonist... the main character... the central figure" for one referent. Fix: pick the clearest term and repeat it; repetition of the right word is not a flaw.
• The treadmill — restating one idea: "In other words,", "Put simply,", "Essentially,". Fix: per sentence ask "what's new here?"; delete rephrasings.
• Reshuffling immunity — paragraphs so self-contained they could be swapped without breaking anything. Fix: make each paragraph need the previous one; merge or cut interchangeable blocks.
• Recap closers — paragraphs ending "Whether you're X or Y...", sections ending "In summary," / "Overall,". Fix: end on the strongest specific point.
• Copula avoidance — "serves as", "stands as", "boasts", "features" dodging plain "is"/"has". Fix: use is/are/has.
• Filler phrases — "in order to" → "to", "due to the fact that" → "because", "has the ability to" → "can", "it is important to note that" → cut.
• Agentless passive — "no configuration is needed", "it is recommended", "changes were made". Fix: name the actor or address the reader.
• Uniform hyphenation — compounds hyphenated even in predicate position ("the report is high-quality"). Humans hyphenate attributively ("a high-quality report") and usually drop it after the noun.

5. Default formatting

Chat-UI formatting habits applied regardless of medium.

• Em and en dashes — in agent-authored text, replace each with (in rough preference order) a period, comma, colon, parentheses, or a restructure. Exception: the author's voice sample or house style uses them deliberately.
• Bold overuse — mechanical emphasis, or erratic 1–4-word bold spans with no shared rule. Keep bold for glossary terms and UI labels.
• Inline-header bullets — **Topic:** sentence lists that should be prose or real structure. Fix: merge into prose, or give items real headings if they earn them.
• Title Case Headings — use sentence case unless house style says otherwise.
• Emoji decoration — 🚀 on headings and bullets. Delete unless the venue expects them.
• Prose-as-table — small tables holding what a sentence would say better. Fix: write the sentence. Tables earn their place only for genuinely 2D data.
• Structural quirks — skipped heading levels (H1 → H3), thematic breaks (---) before every heading. Fix: proper hierarchy, breaks only where a real division exists.
• Curly quotes — weak signal alone (editors auto-curl); normalize to match the document's convention.

6. Machine residue

Literal artifacts of the generation pipeline. Zero judgment required — always remove.

• Citation tokens — oaicite, citeturn0search0, contentReference, oai_citation, attributableIndex, [cite: 1], grok_render_citation_card_json, stray lenticular brackets (【】).
• Placeholders — [Your Name], [INSERT SOURCE URL], 2025-XX-XX, PASTE_URL_HERE, unfilled Mad-Libs blanks. Fill or delete; if you can't fill it, flag it to the author.
• Tracking params — utm_source=chatgpt.com|openai|copilot, referrer=grok.com on URLs. Strip.
• Invisible characters — zero-width spaces/joiners (U+200B, U+200D), soft hyphens (U+00AD), homoglyphs. Normalize to plain text.

Judgment Rules

Clusters convict; single tells don't. Every pattern above occurs in clean human writing. One em dash means nothing; em dashes plus rule-of-three plus "vibrant tapestry" plus a "Conclusion" section is a confession. Flag passages, not words.

Tiered vocabulary. Not every AI-word is equally damning:

• Tier 1 — fix on sight: delve, tapestry (figurative), testament (figurative), underscores (verb), leverage (verb), multifaceted, realm (abstract), interplay, "it's worth noting", "plays a crucial role in", "in today's ... landscape".
• Tier 2 — fix when clustered (2+ per paragraph): ensuring/ensures, highlights, reflects, fosters, showcases, robust, seamless, pivotal, crucial, vibrant, comprehensive, significantly, effectively, "rather than", moreover, furthermore.
• Tier 3 — never flag alone: key, essential, important, significant, various, valuable, notably, such as. Ordinary words; evidence only inside a Tier 1/2 cluster.

Word lists go stale as models change; the mechanism families don't. When a word feels machine-frequent but isn't listed, judge it by mechanism: is it inflating, hedging, or padding?

What NOT to flag — these are not evidence on their own: perfect grammar and polish; formal or academic vocabulary outside the tiers; mixed registers; "bland" prose without specific tells; one short emphatic sentence; common transitions in isolation; curly quotes alone; em dashes alone in human-authored text; unsourced claims; letter-style openings in correspondence.

Preserve the human signals. When you see these, lean toward leaving the prose alone: specific hard-to-fabricate detail; mixed feelings and unresolved tension; era-bound slang and references; genuine asides, parentheticals, and self-corrections; variety the author can defend. Over-editing these destroys what makes text human.

Application

Writing mode (drafting new prose): apply the six families as constraints while composing. State claims directly, prefer specifics you actually have, vary rhythm, format only what earns formatting. Cheaper than editing slop later.

Editing mode (fixing existing text):

1. Read the whole text. One-line voice read before touching anything: " for , register <formal/neutral/casual>."
2. Inventory the substance. List the claims and facts the text carries. This is the conservation checklist for step 5.
3. Sweep by family (1–6 above). Mark clustered passages, not isolated words.
4. Rewrite marked passages. Preserve register, conserve substance, never fabricate specifics.
5. Self-audit. Ask: "What still reads as AI-generated here?" Verify the substance inventory survived. Fix what the audit surfaces.
6. Mechanical pass. Run the residue checks below.
7. Deliver the final text plus a brief summary of what changed and why.

Reviewing mode (verdict without rewrite): sweep by family, report clustered passages with locations and the mechanism behind each, note human signals that argue against AI origin. No verdict from a single tell.

Mechanical residue checks

On files, run these before delivering (adjust to the target's legitimate style):

rg '—|–'                                          # em/en dashes
rg 'oaicite|citeturn|contentReference|oai_citation|attributableIndex|grok_render'
rg 'utm_source=(chatgpt|openai|copilot|perplexity)|referrer=grok'
rg '\[(Your|INSERT|TODO)[ _]|XX-XX|PASTE_|_HERE\b'  # placeholders
rg -P '\x{200B}|\x{200D}|\x{00AD}|\x{FEFF}'          # invisible characters

Any hit is either fixed or consciously kept with a reason (house style, quoted text).

Reference

• Full pattern catalog — ${CLAUDE_SKILL_DIR}/references/patterns.md — per-pattern before/after examples for all six families, extended words-to-watch lists, and source notes (Wikipedia's "Signs of AI writing", 2026 corpus studies). Read when a specific call feels ambiguous or when you need rewrite examples for a pattern.

Critical Rules

• Never fabricate specifics — vivid lies are worse than dull truths.
• Conserve every claim; rewrite, don't summarize.
• Clusters convict; single tells don't. Flag passages, not words.
• Over-editing human voice is as bad as under-editing machine slop.
• Fix the writing for readers, never for detectors.