--- name: humanize description: Transforms polished, AI-generated, or overly formal writing into natural, human-sounding prose. Removes mechanical patterns typical of LLM output — fluffy emphasis, canned phrasing, excessive parallelism, and forced structure. Adds organic roughness appropriate for informal contexts (blogs, comments, emails, fiction, social media). Use when the user wants text to sound less robotic, more natural, less "AI-written," more conversational, or when editing output from language models that feels sterile, formulaic, or overly emphatic. user-invocable: true license: LicenseRef-MutuaL-1.2 metadata: author: Amolith --- # Humanizing Text: Quick Start The goal is not to make text "worse" but to make it _believably human_. Humans are inconsistent, occasionally sloppy, and rarely follow writing-school rules. LLMs are the opposite. Invert LLM defaults. ## Step 1: Strip the AI polish Remove or replace these dead giveaways before anything else: **Overwrought significance phrases:** - "stands as a testament to" → "is" or "shows" or just delete - "a pivotal moment in" → "when" or "this was" - "marking a shift toward" → "then" or "after that" - "underscoring its importance" → delete entirely - "an enduring legacy of" → "people still remember" or delete - "in the broader landscape of" → delete **Canned analysis phrases:** - "highlighting..." / "emphasizing..." / "underscoring..." (as sentence-ending participles) → delete or recast as a standalone sentence - "reflecting broader trends in" → delete - "contributing to the ongoing dialogue" → delete - "valuable insights into" → "ideas about" or delete **Promotional fluff:** - "boasts a vibrant" → "has" - "nestled in the heart of" → "in" - "rich tapestry of" → delete, describe the thing directly - "groundbreaking" → "new" or specific detail - "renowned" → "known for [specific thing]" or just name-drop **Media-coverage chest-beating:** - "has been featured in" / "profiled in" → "was covered in" or just name the publication inline - "independent coverage in notable outlets" → name the outlets or skip - "maintains an active social media presence" → "posts on X" or delete **Avoidance of simple verbs:** - "serves as" → "is" - "stands as" → "is" - "marks a" → "is" - "boasts a" → "has" - "offers a" → "has" or "gives" - "features a diverse array of" → "has" **Negative parallelisms / unearned grandiosity:** - "not just X, but also Y" → "X and Y" or say what you mean directly - "It is not ..., it is ..." → one direct clause - "no ..., no ..., just ..." → one direct clause - "More than just [category]" → describe the thing directly without the elevation pitch **Rule-of-three lists:** - "adjective, adjective, and adjective" → trim to one or two, or vary the structure - "short phrase, short phrase, and short phrase" → break the rhythm: make one longer, merge two, drop one **AI vocabulary (high-density red flags):** - Delete or replace: additionally, delve, crucial, pivotal, robust, intricate, interplay, tapestry, testament, underscore, vibrant, meticulous, fostering, enhancing, enduring, bolster, garner, align with, resonate with, exemplify, showcase **Em dashes:** - Replace with commas, parentheses, or just separate sentences. If you keep them, use `--` (two hyphens) instead of `—`, and use them sparingly. **Bulky structure:** - Kill rigid "Challenges and Future Directions" / "Future Outlook" sections unless the user explicitly asked for them - Kill the formulaic "Despite [positive], [subject] faces challenges..." paragraph structure - Kill the "In conclusion" / "In summary" wrapper - **Arbitrary quantification:** "3 Key Pillars," "5 Core Strategies" — AI forces fluid concepts into numbered baskets. If there aren't exactly three distinct things, don't pretend there are. ## Step 2: Add human irregularities Vary technique by desired register. Use _lightly_ — one or two touches per paragraph, not every sentence. ### Sentence rhythm - **Vary sentence length dramatically.** Drop in a one-word sentence or a 40-word ramble next to a 7-word declarative. - **Start sentences with "And" or "But"** occasionally. - **Use sentence fragments.** "Anyway." "Not really." "The usual." - **Break a long sentence with an aside** that could have been a footnote: "the thing — and this was back in 2019, before everything changed — happened." ### Informal punctuation - Replace some em dashes with double hyphens `--` - Use en-dashes where a human might (sloppily) use hyphens - Let a comma splice slip through if the tone is casual enough - Drop parentheses for appositives instead of em dashes sometimes - Use "..." for trailing off mid-thought - End an occasional sentence with a preposition if that's how people actually talk ### Contractions and compression - Expand contractions for formality. Contract them for informality. Most AI output under-uses contractions. - Use "can't," "won't," "it's," "that's," "there's" in casual text. - Drop auxiliary verbs in casual speech patterns: "Guess so," "Makes sense." ### Imperfections - **Occasional typos or misspellings** in very informal contexts (not for anything important): "recieve," "definately," "seperate" — no more than one per few hundred words - **Minor grammar slips** that don't impede readability: "me and Jake went," "should of," "less people" — only when the register is extremely casual - **Inconsistent capitalization** of words-for-emphasis in informal writing: "it was Bad" - **Elision:** "wanna," "gonna," "kinda," "sorta," "dunno" — appropriate for dialogue, social posts, very casual blogging ### Vocabulary shifts - Replace Latinate abstractions with Anglo-Saxon specifics. - Swap "utilize" for "use," "facilitate" for "help," "leverage" for "use," "implementation" for "doing it." - Swap "individuals" for "people," "purchased" for "bought," "located" for "in." - **False ranges:** "from biology to medieval history" when listing unrelated items — AI loves fake spectra. Say "including biology and medieval history" or just list them. - Use "really," "pretty," "kind of," "basically," "actually" as verbal tics — sparingly, but enough to feel spontaneous. - Drop a colloquialism: "whatever," "I mean," "honestly," "to be fair," "look," "like" (as filler) ### First-person and opinion - AI defaults to third-person neutrality. Add first-person for blogs, reviews, letters: "I thought," "we tried," "I didn't get it at first." - Add a mild opinion or subjective reaction: "which seemed weird," "it sounded better than it was." - **Take a stand instead of hedging.** AI is RLHF-conditioned to do false "on the one hand, on the other hand" balance. Humans pick a side (unless they genuinely feel ambivalent). State what you think. Use hyperbole for effect — exaggeration is human; terrified neutrality is AI. - **If you don't know, say so.** AI papers over gaps with confident filler ("while specific details are limited..."). Humans say "I don't know" or "nobody's sure" or just skip it. ### Specificity over abstraction - AI replaces specific facts with generic praise. Reverse it: "a revolutionary titan of industry" → "the guy who invented the train coupler in 1873." - Name a concrete thing instead of a concept: not "the vibrant cultural landscape" but "the weekly farmer's market and the punk venue that got shut down." ## Step 3: Match the target register The degree of humanization depends on context: | Context | Techniques | | ----------------------- | ----------------------------------------------------------------------------------------------------- | | Blog post / newsletter | Contractions, first person, asides, varied sentence length | | Social media comment | Fragments, elisions, lowercase, occasional typos, emoji | | Business email (casual) | Contractions, lighter structure, drop promotional phrasing, one conversational aside | | Fiction dialogue | Elisions, interruptions, trailing thoughts, imperfect grammar | | Academic-adjacent essay | Kill AI vocabulary, replace vague analysis with specific claims, vary structure, keep grammar correct | | Forum post | All of the above, plus lowercase aggression, sarcasm, or enthusiasm as appropriate | ## What NOT to do - Don't make every sentence sloppy. Irregularity only works next to regularity. - Don't add typos to anything where accuracy matters (legal, medical, technical instructions, citations). - Don't overdo the "uh" and "like" verbal filler — it reads as parody. - Don't strip all structure. Humans use structure; they just don't use _identical_ structure every time. - Don't make the text _harder to read_. Make it _more natural to read_. ## Reference See [references/DETAILED_PATTERNS.md](references/DETAILED_PATTERNS.md) for exhaustive AI-vocabulary lists and model-era-specific tells. See [references/REPLACEMENTS.md](references/REPLACEMENTS.md) for alternative phrasings.