humanize

Transform verbose LLM-generated text into condensed, natural, human-written prose. Two-pass process: rewrite, then self-audit.

For worked before/after examples, see examples.md. For the research grounding (Kobak/Juzek findings, expanded word inventory, the convergence model), see reference.md.

Why all these rules together? AI text is identified by convergence, not any single tell: low lexical diversity + uniform sentence rhythm + overuse of prestige vocabulary + absence of personal idiosyncrasy. A perfect rewrite addresses all four. Removing slop is necessary but not sufficient — you also need to inject specificity (concrete numbers, names, claims) where the original was vague.

Yellow flags — don't over-correct. Some signals look AI-ish but aren't reliable on their own. Sophisticated vocabulary in a professional context (a consultant writing "ascertain the root cause") is just professional vocabulary. Lack of typos likely means Grammarly, not ChatGPT. Lack of contractions might be ESL, formal register, or stylistic choice. Don't strip these unless they cluster with real tells. The skill rewrites slop, not professionalism — when in doubt about a single Tier 1 word in otherwise idiosyncratic human writing, leave it.

Self-reference escape hatch. When the source text is about AI writing (a blog post listing AI tells, a tutorial, a skill file), quoted examples and illustrations are exempt. Text inside quotation marks, code blocks, or explicitly marked as illustrative ("for example, AI might write…") should NOT be rewritten. Only flag patterns in the author's own prose. Without this rule, the skill would mangle every article ever written about AI slop, including its own SKILL.md.

---

Pass 1: Rewrite

Step 1: Cut the filler

Most LLM text is 40-60% filler by volume. Delete before you rephrase.

Substance check first. Before rewriting, ask: does this text make any falsifiable claim? Specific numbers, names, dates, behaviors, mechanisms, comparisons? If after stripping filler there's nothing concrete left — just abstractions about "empowering," "elevating," "transforming," "unlocking value" — don't dress up emptiness in fewer words. Return:

(no substantive content: <one-line description of what the source was, e.g. "marketing copy with no concrete claims">)

The user can override with "rewrite anyway" if they need polished slop for some reason. Default is to call out the void.

Filler phrases — delete on sight:

• "It's worth noting that...", "It's important to understand that...", "It should be noted that..."
• "As previously mentioned...", "As noted above...", "As discussed earlier..."
• "At the end of the day...", "When it comes to...", "In terms of...", "The fact that..."
• "In order to..." → "To..."
• "Due to the fact that..." → "Because..."
• "I think it's fair to say..." → just say it
• "Worth [verb]ing" vague endorsement: "worth reading," "worth a look," "worth your time," "worth checking out." Substitutes a generic thumbs-up for a specific reason. Say why it matters instead.

Hedge stacking — collapse to one claim or delete:

• "may potentially" → "may" or just state it
• "could possibly" → "could" or just state it
• "it is generally considered" → state the thing
• "tends to typically" → pick one

Intensifiers and filler adverbs — delete:

• Certainly, Absolutely, Indeed, Essentially, Basically, Fundamentally
• Very, really, truly, literally, actually, just, quite, rather, simply
• Importantly, Significantly, Notably, Interestingly, Remarkably

Sycophantic openers — delete entirely:

• "Great question!", "That's a really interesting point", "I'd be happy to help with that!"
• Any compliment about the reader's question or thinking
• Forbidden first words: Great, Certainly, Okay, Sure, Absolutely, Of course. Never start a response with these. (SycEval 2025: 58% of LLM responses contain sycophancy with 78.5% persistence once triggered.)

Chatbot artifacts — delete entirely:

• AI self-references that leak into copy: "As a large language model," "my training data," "I am an AI."
• Knowledge-cutoff disclaimers: "As of my last update," "While specific details are limited based on available information," "I don't have access to real-time data." Either find the information or remove the hedge — never publish a sentence that admits the writer didn't look something up.
• Conversational closers from chat interfaces: "I hope this helps!", "Let me know if you'd like me to expand on any section," "Feel free to reach out."

Meta-commentary — delete:

• "Let me break this down:", "Here's what you need to know:"
• "In this section, we'll explore...", "Let's take a closer look at..."
• "Let's dive in," "Let's explore," "Let's break this down" — any "Let's + verb" used as a transition rather than a real invitation.
• "There are several key factors to consider: First... Second... Third..."
• Don't announce what you're about to say. Just say it.

Acknowledgment loops — delete:

• "You're asking about…", "To answer your question…", "The question of whether…"
• AI restates the prompt before answering. The reader knows what they asked. Just answer.
• Same goes for opening a section by summarizing the previous one. Trust the structure.

Reasoning chain artifacts — delete:

• "Let me think step by step," "First, let's consider," "To approach this systematically," "Step 1:," "Working through this logically."
• Chain-of-thought scaffolding leaking into prose. The reader doesn't need to see the model's internal monologue. State the conclusion, then the evidence.

Numbered list inflation — flag:

• "Three key takeaways," "Five things to know," "Top 7," "Here are the top ten."
• AI defaults to numbered lists because they're structurally safe. Only use numbered lists when the content has genuinely that many discrete, parallel items. If you're padding to hit a number, the list shouldn't exist.

Unearned profundity — delete:

• "Something shifted." "Everything changed." "But here's the thing." Narrative-pivot phrases that promise revelation and deliver none.
• Mid-sentence rhetorical questions: "But now?" "The solution?" "What does this mean?" — usually set up so the AI can answer itself in the next clause.
• Inspirational closers: "Whatever your X is — start strumming." "Now go and do." "Here's to the ones who…" AI loves ending on a beat. Cut the closer or replace with a concrete next step.
• Closing teasers: "I could go deeper if you want," "Let me know if you'd like more on X." Manipulative engagement-bait suffixes from chat-trained models. Cut.
• Magic adverbs as profundity injectors: "quietly," "gently," "subtly" used to add false depth ("This quietly changed everything"). Either show the quietness/subtlety with specifics, or cut the adverb.
• Invented concept labels: "the supervision paradox," "the X principle," "the Y problem nobody's naming." Labeling mundane observations as named paradoxes/principles to sound profound. Either cite a real source for the term or describe the thing without giving it a fake name.
• Pseudo-punchlines: "The result:", "The point is:", "The bottom line:", "The truth is:", "Here's the thing:" — false-punchline framing that promises a payoff and delivers an ordinary statement. Either earn the punchline with setup or just state the claim.
• Persuasive authority tropes: "the real question is," "at its core," "in reality," "fundamentally," "the deeper issue," "the heart of the matter," "what really matters." LLMs use these to pretend to cut through noise to deeper truth — but the next sentence usually just restates an ordinary point with extra ceremony. Cut the framing; keep the claim.
• Futuristic/scenario invitations: "Imagine a world where…", "In a world where…", "Picture this:" — fiction-trailer openers used to dress up ordinary observations. Cut or replace with a concrete present-tense example.

Step 2: Kill the LLM vocabulary

Three tiers — a flat ban list creates false positives.

Tier 1: Always flag and replace. These words appear at 5–25× above human baseline frequency in LLM output (Kobak et al., Science Advances, 2025). One occurrence is a tell.

Kill	Replace with
delve	(delete, or: examine, look at, dig into)
leverage	use
utilize	use
facilitate	help, enable, let
comprehensive	(delete, or be specific about scope)
robust	(delete, or name the specific quality: fast, reliable, tested)
streamline	simplify, speed up
holistic	(delete)
synergy	(delete)
multifaceted	(delete)
showcasing	showing
underscores	shows, highlights
tapestry	(delete — unless literally about fabric)
realm	area, field, domain
beacon	(delete)
cornerstone	(delete, or: foundation, basis)
nexus	(delete)
interplay	interaction, relationship
paradigm	model, approach
testament	proof, evidence
pivotal	key, critical, important
transformative	(delete, or prove the transformation)
unprecedented	(delete, or prove it's actually unprecedented)
bespoke	custom
myriad	many
embark	start, begin
encompass	include, cover, span
elucidate	explain, clarify
juxtapose	compare, contrast
commence	start, begin
ascertain	find out, determine
boasts	has
serves as	is
despite challenges...continues to thrive	(name the challenge and the response, or cut)
the future looks bright	(cut — say something specific or nothing)
only time will tell	(cut — say something specific or nothing)
best practices	what works, proven methods
thought leadership / thought leader	expert, authority (or describe their actual contribution)
key takeaways	findings, conclusions, points
actionable insights	(cut, or name the action)
dive deep / take a deep dive	look at, examine
unveil	show, reveal
shed light on	clarify, explain
in a world where / imagine a world where	(cut — describe the present-tense reality)
as a large language model / my training data	(cut entirely — AI self-reference leak)
esteemed / enlightening	respected, useful
crafting	writing, building, making
skyrocketing / skyrocket	grew sharply, jumped
surpassing	exceeding, beating
garnered	earned, attracted, received
advancements	improvements, progress
aligns with	matches, fits
comprehending	understanding
intricacies	details, complications

Tier 2: Flag when they cluster. One use might be fine. Two in a paragraph is a pattern. Three is slop.

landscape, ecosystem, innovative, seamless, cutting-edge, groundbreaking, game-changer, empower, harness, unleash, supercharge, foster, navigate, unlock, elevate, amplify, spearhead, catalyze, nuanced, paramount, invaluable, meticulous, intricate, ever-evolving, plethora, treasure trove, kaleidoscope, enhance, optimize, augment, reimagine, revolutionize, underpin, poised, burgeoning, nascent, quintessential, paradigm-shifting, profound, incredible, amazing, fantastic, brilliant, stunning, epic, exciting, remarkable.

Tier 3: Flag only at high density. Common words that are individually fine but become tells when they pile up. Threshold: three or more in a single paragraph, or one per ~75 words across longer text. Below that, leave them alone.

furthermore, moreover, notably, additionally, consequently, accordingly, nevertheless, ultimately, thereby, whilst, particularly, across (overused as transition), within (overused as transition), exhibited, journey (metaphorical), gateway, milestone, linchpin, straightforward, genuinely, honestly, resonate, illuminate.

Tier 3a: Dramatic words and absolutes (flag when hyperbolic). These get reached for when AI wants to inject false stakes. Fine when literally accurate; problematic when stretched.

• Dramatic words: paradox, trap, illusion, magic, destiny, chaos, crisis, explosion, apocalypse, miracle, revolution.
• Absolute terms used as filler: always, never, everything, nothing, completely, entirely, totally — flag when the claim is clearly not absolute. "Always" in "always works" is a tell; "always" in "this function always returns true" is correct.

Step 3: Fix structural patterns

Equally diagnostic as word choice. AI text has structural fingerprints.

• Em dashes — limit to one per 500 words max. Strongest punctuation-level AI tell. Replace with periods, commas, colons, or parentheses.
• Negative parallelisms — rewrite or delete. "It's not just X — it's Y" or "Not merely X, but Y" appears in nearly every LLM's default output. Just say Y.
• Break the rule of three. LLMs default to triplets ("innovation, inspiration, and insights"). Use two items or four. Or pick the one that matters.
• False agency — remove. Don't give inanimate things human verbs. "This framework enables teams to..." → "Teams can use this framework to..." or just "Teams can...". Objects don't enable, empower, unlock, or drive. People do things.
• Copula inflation — use "is". "Serves as", "acts as", "functions as", "operates as" → "is". If something is a thing, say it is that thing.
• Significance inflation — deflate. "Marking a pivotal moment in the evolution of..." → state the fact. "Representing a watershed shift in..." → say what happened.
• Rigid scaffolding — remove. Don't reproduce intro-body-conclusion for anything under 1000 words. Start with the most important point. End when you're done.
• Random bolding / italics — strip. AI bolds words that don't carry the load of the sentence. Remove bolding unless the bolded phrase is genuinely the central claim. Same for italics.
• Unicode formatting tricks — convert to plain text. 𝗯𝗼𝗹𝗱, 𝘪𝘵𝘢𝘭𝘪𝘤, →, ×, ✅/📊/💡 leading bullets. Use real markdown if the surface supports it; otherwise plain ASCII. Decorative Unicode is an AI tell in 99% of professional contexts.
• Generic metaphors — cut or anchor. AI metaphors are plausible but ungrounded ("teaching your fingers to dance," "a puzzle piece clicking into place"). Human metaphors are either highly specific (personal experience: "like the time I…") or culturally resonant (shared reference). If a metaphor doesn't anchor in either, delete it. Don't replace with another generic one.
• Promotional language — replace with plain description. "Nestled within breathtaking foothills," "a vibrant hub of innovation," "a thriving ecosystem." Tourism-brochure prose. Replace with what the thing actually is: "is a town in the Gonder region," "has 12 startups."
• False ranges — cut or list real items. "From the Big Bang to dark matter," "from ancient civilizations to modern startups," "from healthcare to fintech." Pairs of unrelated extremes that sound sweeping but commit to nothing. List the actual topics covered, or pick the one that matters.
• False concession structure — pick a side or specify. "While X is impressive, Y remains a challenge." "Although the team has made strides, gaps remain." Performative balance with vague halves. Either name the impressive specific and the challenge specific, or pick a side and argue it.
• Parenthetical hedging — commit or cut. "(and, increasingly, Z)," "(or, more precisely, Y)," "(and perhaps more importantly, W)." Asides that sound nuanced without committing. If the aside matters, give it its own sentence. If not, cut it.
• Novelty inflation — describe the contribution, don't claim invention. "He introduced a term," "She coined the phrase," "a failure mode nobody's naming," "the insight everyone's missing." Most ideas are applications of existing concepts, not inventions. Factually risky AND promotional. Replace with what the person actually did with the concept, not that they discovered it.
• Emotional flatline — show, don't claim. "What surprised me most…", "I was fascinated to discover…", "What struck me was…", "The most interesting part…", "It hit different." AI claims emotion as structural crutch without conveying it through the writing. If the thing is genuinely surprising, the reader should feel that from the content. Cut the emotional claim or earn it with detail.
• Inline-header lists — strip the bold leads. Lists where each item starts with a bolded keyword that repeats: "Performance: Performance improved by…" / "Speed: Speed matters because…" Strip the bold header. If the items need headers, they should probably be paragraphs.
• Title case in subheadings — use sentence case. "Strategic Negotiations And Key Partnerships" → "Strategic negotiations and key partnerships." Title case is for the piece's main title at most, not every H2/H3.
• Curly quotation marks — convert to straight. ChatGPT defaults to "curly" quotes (left and right) instead of "straight" ASCII quotes ("..."). Convert to straight quotes — they don't break copy-paste into code or terminals, and humans typing don't reach for them.
• Tailing negation fragments — write as real clauses. "The options come from the selected item, no guessing." "It just works, no wasted motion." Verbless tag-on negations are a tell. Either fold into the sentence as a real clause ("...without forcing the user to guess") or delete the fragment.
• Hyphenated word-pair overuse — strip the hyphens or vary. AI hyphenates compounds with eerie consistency: "third-party," "cross-functional," "client-facing," "data-driven," "decision-making," "well-known," "real-time," "long-term," "end-to-end," "high-quality." Humans are inconsistent — some hyphenate, some don't, almost never uniformly. If a piece hyphenates every common compound, strip most of them. Keep technical compounds where the hyphen disambiguates.
• Fragmented headers — drop the warm-up line. A heading followed by a single sentence that just restates the heading before the real content begins. Before: "## Performance / Speed matters. / When users hit a slow page..." After: "## Performance / When users hit a slow page..." The restatement adds nothing.
• Dramatic countdown negation — collapse into a real claim. "Not X. Not Y. Just Z." Distinct from "not just X, but Y" — this is the stair-step variant that builds false tension across multiple short sentences. Replace with a single direct statement of Z, or one short negative clause if the contrast actually matters.
• Patronizing analogies — cut "Think of it as…". "Think of it as a Swiss Army knife for X." "Imagine if your laptop had a personal assistant." Assumes the reader needs simplification they didn't ask for. Either trust the reader or use a precise term — don't substitute a folksy stand-in.
• Historical analogy stacking — pick one or none. "Apple didn't build Uber. Facebook didn't build Spotify. Airbnb didn't build…" Rapid-fire company examples chained for false-authority effect. Pick the single analogy that actually fits the argument, or cut all of them and make the point directly.

Step 4: Fix the rhythm

Sentence-length variance. AI text has measurably uniform sentence length. Human text is bursty — short punches mixed with longer constructions.

• Vary sentence length deliberately. After two medium sentences, use a short one. Then maybe a longer one.
• Rough mix: 20% short (under 8 words), 50% medium (8–20 words), 30% long (20+ words). Don't measure — feel for monotony and break it.
• If every sentence in a paragraph is 15–20 words, something is wrong.
• One- or two-word sentences are fine occasionally. They punch.
• Don't start three consecutive sentences with the same word or structure.

POV and tense variance. A separate phenomenon from sentence rhythm. AI picks a person (first / second / third) and stays uniform across a whole piece. Humans switch when rhetoric calls for it ("you do X, but I noticed Y"). Don't force a switch, but don't lock in either — uniform POV across long text is itself a tell.

Step 5: Apply the rewrite principles

1. Active voice. Subject-verb-object. Name the actor. "The decision was made to..." → "We decided to..."
2. One idea per sentence. Split when "and" joins two separate thoughts. Keep when "and" enumerates within a single thought ("I tested it locally and on staging" stays; "I tested it and we should ship it" splits).
3. Preserve technical accuracy. Terse ≠ vague. Keep specific numbers, names, versions, constraints. Cut the fluff around them, not the facts.
4. Match the register. Technical docs stay technical but tight. Casual stays casual and tight. Don't flatten everything to one voice.
5. Target 40–60% length reduction. Under 20% means you didn't cut enough. Over 80% means you lost content. Exception: the substance-check void-flag (Step 1) overrides this — if there's nothing concrete to compress, return the void flag, not laundered emptiness.
6. Inject specificity where the original was vague. Cutting slop is necessary but not sufficient. Replace "issues" with "errors and latency spikes." Replace "significant growth" with "30% YoY." Replace "leading provider" with the actual name. When the source has SOME concrete content but vague spots: leave the gaps visible rather than invent — flag the vague nouns as slop signals but don't fabricate replacements. When the source has NO concrete content at all: fall back to the substance-check void flag from Step 1.
7. Prefer numeric constraints over adjectives. "Three sentences" beats "concise." "Under 100 words" beats "tight." When you instruct yourself or others, use numbers.
8. Patch vs. rewrite from scratch. If the source has 5+ Tier 1 hits across multiple categories, 3+ structural patterns triggered, AND uniform sentence/paragraph rhythm, patching individual phrases won't fix it — the structure itself is AI-generated. State the core point in one sentence, then rebuild from there. Don't try to surgically humanize unsalvageable slop; rewrite it.
9. Never invent. Do not fabricate personal anecdotes, named people, citations, statistics, or examples that weren't in the source. If the source claims "studies show" without naming the study, do not invent a plausible one. Either flag the vague attribution and leave it visible, or cut the claim. Inventing details to humanize is worse than the slop you replaced — it manufactures false authority. Hypothetical examples are allowed only if explicitly framed as hypothetical ("imagine," "suppose").
10. Cut quotables. If a sentence reads like a pull-quote — short, polished, tweetable, designed to be the line in a magazine pull-out — rewrite it. AI defaults to manufacturing aphorisms because RLHF rewards them. Real prose has connective tissue between strong sentences. If everything sounds quotable, nothing does.

---

Pass 2: Self-audit

After rewriting, review against this checklist. If any item fails, fix it before returning.

1. Tier 1 words. Search for every Tier 1 word. If any survived, replace them.
2. Em dashes. Count them. More than one per 500 words? Replace extras.
3. "Not just X, but Y". Any negative parallelisms? Rewrite as direct statements.
4. Triplets. Any rule-of-three patterns? Break them.
5. False agency. Inanimate things doing human actions? Fix.
6. Copula inflation. Any "serves as", "acts as", "functions as"? Use "is".
7. Sentence rhythm. All sentences roughly the same length? Vary them.
8. Meta-commentary. Any "In this section...", "As we'll see..."? Delete.
9. Sycophancy. Any leftover compliments, "great point", or similar? Delete.
10. Lexical diversity. Are the same prestige words recurring across the rewrite? AI text has measurably lower type-token ratios. Vary word choice across the whole output. Caveat: don't synonym-swap within a tight passage just to avoid a repeat — that's the "elegant variation" tic (also AI). Real writers repeat the right word when it's the right word; they vary at the document level, not the sentence level.
11. Specificity check. Did you replace at least some vague nouns ("issues," "factors," "things") with concrete ones? If everything is still abstract, the rewrite is incomplete.
12. Self-critique question. Read the output back and ask: "What would make this obviously AI-generated?" If you can answer, fix that thing. Repeat until you can't.
13. Substance check (re-run). Does the polished output actually claim anything? If you stripped slop and ended up with smoothly-written nothing, fall back to the empty-content fallback from Step 1. Don't ship laundered emptiness.

---

Output format

Return only the rewritten text. No preamble ("Here's the condensed version:"), no postamble ("Let me know if you'd like adjustments!"), no meta-commentary. Just the text.

If the user pastes multiple distinct blocks, rewrite each separately and return them in order. If they ask for options, return 2–3 numbered variants, no commentary between them.