# Humanize > Strip AI writing tells out of text so it reads like a person wrote it. Use whenever the user says "Humanize", "humanize this", "make this sound human", "does this sound like AI", "de-AI this", "clean up the AI voice", or asks to remove AI tells / AI-isms / robot voice from a draft, resume, cover letter, email, post, or any other writing. Also use proactively when the user pushes back that something reads generic, stiff, corporate, or "like ChatGPT wrote it." - Skill: `vyquemendoza/humanize` (Agent Skill) - Install (CLI): `npx skillmds@latest add vyquemendoza/humanize` - Raw SKILL.md: https://api.skillmd.com/api/skills/vyquemendoza/humanize/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Productivity - Author: vyquemendoza (https://skillmd.com/u/vyquemendoza) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/vyquemendoza/humanize --- # Humanize Rewrite text so it stops reading as machine-generated. The checklist below comes from Wikipedia's *Signs of AI writing* field guide (WP:AITELLS), filtered to prose and style tells, plus the user's own blocklist. **Source last checked: 2026-08-03.** A daily task re-reads the page and updates this skill when the list changes. ## How to run **Default scope:** the most recent thing you wrote in this conversation. Rewrite it in place, in full. **If the user names a file** (resume, cover letter, .docx, .md), edit that file instead. Preserve its formatting and structure. **If the user pastes text**, rewrite the pasted text. **Output:** the clean version only. No change log, no "here's what I fixed," no before/after table, unless the user asks. They want the fixed writing, not a report about the fixing. ## The core problem you're solving AI writing regresses to the mean. It swaps specific, odd, load-bearing detail for generic praise that could apply to anything. "Invented the first train-coupling device" becomes "a revolutionary titan of industry." The writing gets simultaneously vaguer and louder. So the fix is almost never *delete the flagged word*. It's *restore the specific thing the flagged word was covering for.* If you strip "showcasing her commitment to innovation" and leave a hole, you've made the sentence shorter and equally empty. Ask what actually happened, and if the source text says, put that in. If it doesn't, cut the clause entirely rather than swapping one vague phrase for another. This is the single most common way a humanize pass fails: it becomes a find-and-replace on a banlist, the tells disappear, and the writing still reads synthetic because the underlying vagueness never moved. ## Content tells **Puffed-up significance and legacy.** Claims that some arbitrary detail represents a broader trend or marks a turning point. > stands as / serves as, is a testament to, is a reminder that, played a crucial/pivotal/vital/key role, underscores its importance, reflects broader, symbolizing its enduring/lasting/ongoing, contributing to the, setting the stage for, marking a shift, represents a shift, key turning point, evolving landscape, focal point, indelible mark, deeply rooted Cut these. If the significance is real, it should be visible from the facts without narration. **Superficial analysis bolted onto the end of sentences.** Usually an "-ing" participle phrase that adds interpretation nobody asked for. > ...highlighting her leadership. ...ensuring seamless delivery. ...reflecting the team's values. ...contributing to overall growth. ...fostering collaboration. ...enhancing engagement. The sentence almost always ends better one clause earlier. Delete the participle tail. **Promotional / travel-brochure register.** Even when the brief is neutral, the output drifts toward advertising. > boasts a, vibrant, rich, profound, enhancing, showcasing, exemplifies, commitment to, natural beauty, nestled, in the heart of, groundbreaking, renowned, featuring, diverse array, seamless, robust, cutting-edge, world-class **Vague attribution / weasel wording.** Attributing opinion to an authority that is never named, or inflating one source into many. > Industry reports suggest, Observers have noted, Experts argue, Some critics say, Many have pointed out, several sources, it is widely believed Name the source or drop the claim. **Canned notability stacking.** Piling up proof that something matters instead of just showing it. > received independent coverage, was profiled in, featured in national media, trade publications, maintains an active social media presence, written by a leading expert **The challenges-and-future-outlook formula.** "Despite its [good things], X faces several challenges..." followed by a hopeful closer, or a "Looking ahead" / "Future Outlook" section. The tell is the rigid shape, not the mention of difficulty. Real challenges written in a real voice are fine. **Knowledge-cutoff and missing-source hedging.** "While specific details are limited," "not widely documented," "based on available information," "as of my last update." Also the personal-life version: "keeps a low profile," "maintains privacy about." Either state the fact or leave it out. **Leftover chatbot chatter.** "Certainly!", "I hope this helps", "Let me know if", "Would you like me to", "Here is a detailed breakdown", "In this section we will discuss." Never survives into a deliverable. **Unfilled placeholders.** `[Company Name]`, `[insert metric]`, `2025-xx-xx`. Scan for brackets before delivering. ## Language and rhythm tells **AI vocabulary clustering.** One of these is coincidence. Five is a fingerprint. They travel in packs. > additionally (especially opening a sentence), align with, boasts, bolstered, crucial, delve, emphasizing, enduring, enhance, fostering, garner, highlight (as a verb), interplay, intricate/intricacies, key (as an adjective), landscape (figurative), meticulous, pivotal, robust, showcase, tapestry (figurative), testament, underscore (as a verb), valuable, vibrant Density is what matters. Judge the paragraph, not the word. "Crucial" once in a page is normal English; three times in a page with "pivotal" and "underscore" nearby is a giveaway. **Copula avoidance.** Machines dodge plain *is* and *are*, reaching for verbs that sound weightier. > serves as → is > stands as → is > functions as / operates as → is > represents → is > features / offers / maintains / boasts → has > refers to → is > began her career as → was Plain *is* is a strength, not a weakness. Put it back. **Negative parallelism.** Setting up a misconception nobody held so the sentence can knock it down. > Not only X, but Y > It's not just X, it's Y > Not X, but Y > No X. No Y. Just Z. > X rather than Y (as a rhetorical flourish) Say the positive claim by itself. **Rule of three.** Three adjectives, three clauses, three examples, over and over. Used to make thin analysis feel thorough. Break the pattern: two items, or four, or one specific one. Vary it across the piece so no rhythm dominates. **Elegant variation.** Compulsively swapping in synonyms to avoid repeating a word, because of the model's repetition penalty. If the subject is a *report*, call it a report every time. Don't cycle through document / analysis / deliverable / piece. **Uniform sentence length.** AI prose runs at one steady tempo. Human writing lurches. Mix a fragment against a long one. Let some sentences run short. ## Formatting tells - **Title Case In Headings** → sentence case. - **Boldface sprayed for emphasis**, especially every instance of a key term, or "key takeaways" style. Bold sparingly, or not at all. - **Inline-header bullet lists**: `- **Thing:** description` repeated down the page. Convert to prose where it reads fine as prose. - **Em dashes.** The most notorious tell, especially spaced ones. This user's hard rule: **at most one em dash per document**, and zero is better. Replace with commas, colons, parentheses, or a period. - **Curly quotes and apostrophes** where straight ones belong, or a mix of both in one document. - **Emoji as decoration** on headings or bullets. - **Thematic breaks (`---`) before every heading.** - **Tiny unnecessary tables** for content that's two sentences of prose. ## The user's personal blocklist These are banned outright, on top of everything above: - "But here's the thing:" - "Then I realized:" - "The result?" - "Hot take:" - "but here's what nobody's saying" - "fluff" - "delve" - "It's not just [X]. It's also [Y]." - "No [X]. No [Y]. Just [Z]." - "From X to Y:" as an opener Her stated preference: punchy and attention-getting, not bland or generic. So when you cut a flourish, don't leave beige in its place. Cut to something with a sharper edge. ## What not to do **Don't sand off the voice.** Removing every rhetorical move leaves flat, affectless prose, which is its own kind of machine-written. Contractions, sentence fragments, a joke, a strong verb, an opinion stated flatly, a specific weird detail: these are what human writing has that AI writing doesn't. Add them, don't just subtract. **Don't treat the list as law.** Wikipedia is explicit that these are descriptive observations, not rules, and that plenty of human writing contains them. "Crucial" in a sentence where the thing genuinely is crucial should stay. A rule of three that lands should stay. Judgment beats compliance. **Don't change the facts.** Numbers, dates, names, titles, claims about what the person did: all stay exactly as they were. You are changing how it reads, not what it says. This matters most on resumes and cover letters, where an invented detail is a real problem. **Don't over-hedge the result.** Some tells are unavoidable in certain genres. A resume bullet will use strong verbs. A cover letter will state significance. Calibrate to the document type instead of applying maximum severity everywhere. ## Quick self-check before delivering Read the rewrite out loud in your head and ask: 1. Could this paragraph describe a different company, person, or project with three words swapped? If yes, it's still generic. Put the specific thing back. 2. Does every sentence run about the same length? 3. How many em dashes? Should be zero or one. 4. Did anything get shorter but not sharper? 5. Does it still sound like the person whose name is on it?