humanize-text-skill
Audit and rewrite Chinese or English content to remove AI tone, then pull it toward a target human voice. Use this skill when asked to remove AI tone, sound human, rewrite naturally, make a draft feel less templated, or match a target voice. Supports detect-only and edit-in-place modes, scene packs, protected spans, and voice profiles.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/lynote-ai--humanize-text-skill)HTML
<a href="/listings/lynote-ai--humanize-text-skill"><img src="/api/listings/lynote-ai--humanize-text-skill/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
git clone https://github.com/fendouai/humanize-text-skill.git cd humanize-text-skill npm test
This repository is usually used as a skill package inside an agent runtime.
- Claude Code: copy `SKILL.md`, `references/`, and `policy/` into `.claude/skills/humanize-text-skill/` - Cursor: use [`cursor-rules/humanize-text-skill.mdc`](./cursor-rules/humanize-text-skill.mdc) - Codex / OpenClaw / Hermes: point the tool at the repo root
ClawHub publish readiness
This repository is prepared for ClawHub upload:
- canonical entry file: `SKILL.md` - runtime support files: `detector/`, `references/`, `policy/` - publish filter: [`.clawhubignore`](./.clawhubignore)
Suggested publish command:
clawhub skill publish . \ --slug humanize-text-skill \ --name "humanize-text-skill" \ --version 0.1.0
`policy/` is a required runtime dependency. If you install the skill manually and omit that directory, `voiceMode` will fail.
Install health check
node -e "const H=require('./detector/patterns.js'); const r=H.analyzeText('值得注意的是,我们打造了一套方案。', { voiceMode: 'casual' }); console.log({ score: r.score, hasVoice: !!r.voice, drift: r.voice && r.voice.drift });"Expected:
- `score` is greater than `0` - `hasVoice` is `true` - `drift` is a number
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
SKILL.md · Manifest · 167 lines
View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)
- ---
- name: humanize-text-skill
- description: Audit and rewrite Chinese or English content to remove AI tone, then pull it toward a target human voice. Use this skill when asked to remove AI tone, sound human, rewrite naturally, make a draft feel less templated, or match a target voice. Supports detect-only and edit-in-place modes, scene packs, protected spans, and voice profiles.
- version: 0.1.0
- compatibility: Any AI coding assistant that supports agentskills.io SKILL.md format (Claude Code, Cursor, VS Code Copilot, Hermes, OpenHands) or OpenClaw. No external tools or APIs required.
- metadata:
- tags: writing editing voice quality bilingual
- agentskills_spec: "1.0"
- openclaw:
- skillKey: humanize-text-skill
- emoji: "\u270D\uFE0F"
- homepage: https://github.com/fendouai/humanize-text-skill
- ---
- # humanize-text-skill
- > Subtraction plus addition. First remove AI-shaped prose, then pull the result toward a target human voice.
- > This skill is bilingual in behavior, but the documentation and operating contract are English-first.
- ## What this skill is and is not
- This is a writing-quality tool, not a verdict. The patterns flagged here are statistically more common in LLM output, but humans under deadline pressure, working in a second language, or drafting in an unfamiliar genre can produce the same shapes. Treat the findings as signals, not proof.
- ## Modes
- - **`rewrite`**: flag AI tone, rewrite the text, and pull toward a target voice if one is set
- - **`detect`**: flag issues only, with no rewrite
- - **`edit`**: edit a file in place with minimal targeted changes
- Every mode runs protected-span detection first so version numbers, commands, paths, errors, quotes, and other anchored facts do not drift.
- ## Working Style
- Work like an editor, not a slogan filter. The goal is not just to delete AI-looking phrases. The goal is to leave the user with text they can actually send.
- Default flow:
- 1. fence protected spans first
- 2. name the dominant problem
- 3. choose the lightest effective move
- 4. run one quick residue pass
- 5. return something usable for the scene
- Default principles:
- - fidelity before smoothness
- - name the problem before changing the sentence
- - use the smallest stable edit that solves it
- - do not invent facts, sourcing, or attitude just to sound more human
- - be direct in direct scenes and conservative in risky scenes
- ## Typical Invocation Shapes
- ### 1. Direct skill call
- ```text
- /humanize-text-skill Please rewrite this so it sounds less like AI:
- [paste text]
- ```
- ### 2. Natural-language request
- ```text
- Use humanize-text-skill to rewrite this README intro so it sounds natural:
- This project serves as a testament to our commitment to innovation...
- ```
- ### 3. File-based request
- ```text
- /humanize-text-skill Please humanize the copy in article.md.
- ```
- ### 4. Detect-only request
- ```text
- /humanize-text-skill Detect mode: tell me what still sounds AI-generated, but do not rewrite yet.
- ```
- ### 5. Voice-targeted request
- ```text
- /humanize-text-skill Rewrite this in a blunt voice for an issue reply.
- ```
- ### 6. Custom voice calibration
- ```text
- /humanize-text-skill
- Here is a sample of my writing:
- [paste 2-3 paragraphs]
- Now rewrite this in my voice:
- [paste text]
- ```
- When a user provides a personal sample, treat it as a custom voice profile:
- - extract rhythm, sentence-length spread, connector habits, and first-person tendency
- - do not copy factual content or opinions from the sample
- - do not violate protected spans just to lower `voice.drift`
- When the user's request is underspecified, prefer these defaults:
- - pasted text -> `rewrite`
- - "take a look" or "check this" -> `detect`
- - "only touch this file/comment/paragraph" -> `edit`
- - README, release note, issue reply, or forum post -> load the matching scene pack
- ## What to remove or fix
- > The categories below are the human-facing rule catalog. Each `###` entry maps to detector coverage, model judgment, or both. The English description is canonical for the contract; bilingual engine behavior still applies.
- ### Tier 1 vocabulary (always flag)
- Words and phrases that are dramatically overrepresented in AI text. Replace them on sight. Chinese examples include opener cliches, abstract business jargon, and social-platform sales talk. English examples include `delve`, `tapestry`, `leverage`, `seamless`, `robust`, `comprehensive`, `game-changer`, `serves as`, and `at its core`.
- ### Tier 2 vocabulary (flag in clusters)
- These words are individually acceptable, but clustering is the signal. In short paragraphs, two or more often means the writing is drifting toward template prose. Keep the best-fit instance and rewrite the rest.
- ### Tier 3 vocabulary (flag by density)
- Common abstract words should only be flagged when they saturate the document. Replace some of them with specifics such as numbers, concrete actions, names, dates, or examples.
- ### Structural anti-patterns (cross-lingual)
- Watch for summary closers, mechanical ordering, binary contrast framing, symmetry padding, empty balance, and other structures that sound manufactured rather than written.
- ### Translation tone (Chinese-specific)
- Chinese drafts can inherit English-thinking structures: stacked passives, long attributive chains, `based on`-style openers, and `through ... to ...` constructions. Rewrite the sentence around natural Chinese syntax instead of patching the surface.
- ### Chatbot artifacts
- Remove assistant residue entirely: greeting fluff, forced enthusiasm, help-closing lines, over-polite acknowledgments, and reasoning-process narration such as `Let me think step by step`.
- ### Significance inflation
- Cut claims that overstate the meaning of ordinary events. State what happened and let the reader decide whether it is pivotal.
- ### Vague attribution
- Phrases like `experts believe` and `research suggests` need a specific source. If there is no source, either cite one or remove the attribution.
- ### False-concession structure
- Balanced-sounding `while X, Y` or `not X, but Y` framing is often used to sound thoughtful without saying much. Make both sides concrete or choose the stronger point.
- ### Promotional language
- Remove ad copy, inflated product language, and corporate uplift. If the sentence would sound strange in a normal conversation, flatten it.
- ### Social endorsement closers
- Phrases such as `worth your time`, `thank me later`, or generic bookmarking prompts usually add no information. Replace them with who the piece is for and why.
- ### Hedge-stacked predictions
- Modal verbs plus stacked hedges (`could potentially`, `may eventually`) cancel each other out. Keep one hedge when uncertainty is real.
- ### Formulaic openers
- Avoid generic scene-setting such as `in today's rapidly evolving world`. Lead with the actual news or claim, then add context if needed.
- ### Emotional flatline
- Statements like `what surprised me most` or `this was deeply meaningful` often name a feeling without earning it. Show the specifics or cut the emotion claim.
- ### Novelty inflation
- Be suspicious of `nobody is talking about this` framing. Unless novelty is clearly supported, frame the idea as one interpretation rather than a revelation.
- ### AI-tool fingerprints (placeholders / citations / UTM)
- Strip mechanical artifacts such as unfilled placeholders, leaked chatbot citation markup, and URL parameters tied to chat tools. These are tool residue, not style.
- ### Rhythm & uniformity (stylometric)
- Even when vocabulary looks fine, drafts can still feel synthetic because the rhythm is too smooth. Watch for sentence-length uniformity, repetitive punctuation behavior, and low variation in grammatical movement.
README.md · README · 262 lines
View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)
- <div align="center">
- # humanize-text-skill
- ### Make AI-written text sound human, not just less AI.
- **Bilingual zh/en engine · zero dependencies · detector + rewrite skill + voice pull**
- </div>
- `humanize-text-skill` is a bilingual writing skill for removing AI-shaped prose and pulling the result toward a target human voice.
- It does two jobs:
- - **Subtraction**: remove AI-shaped vocabulary, structure, rhythm, and chatbot residue.
- - **Addition**: pull the cleaned text toward a target voice such as `casual`, `professional`, or `technical`.
- Unlike simple AI-writing detectors, this project is built as a full skill:
- - it detects issues
- - it decides what is safe to rewrite
- - it protects facts, commands, paths, and quoted text
- - it shapes the final output for scenes like README intros, release notes, issue replies, and status updates
- ## At A Glance
- 
- ## Why It Exists
- Most tools stop at "less AI." `humanize-text-skill` aims for "sounds like a real person wrote this."
- Common targets:
- - Chinese filler and black-box jargon such as `值得注意的是`, `赋能`, `闭环`, and `综上所述`
- - English AI staples such as `delve into`, `seamless`, `serves as a testament`, and `Let's dive in`
- - over-even sentence rhythm
- - false significance and vague authority claims
- - pasted chatbot residue
- ## What It Can Do
- - Detect 50 executable issue types across Chinese and English.
- - Keep the same concept mapped to the same detector type across both languages where symmetry is intended.
- - Route behavior by scene: `chat`, `status`, `docs`, `public-writing`.
- - Support three modes: `rewrite`, `detect`, `edit`.
- - Protect spans that must not drift: numbers, commands, paths, versions, names, logs, and quotes.
- - Compute `voice.drift` separately from `score`, with concrete pull suggestions.
- - Document **17 pattern categories** in the skill contract.
- - Lock examples and engine behavior with `143 CI checks`.
- ## Quick Example
- ```bash
- node -e "const H=require('./detector/patterns.js'); console.log(H.analyzeText('值得注意的是,我们打造了一套赋能开发者社区的方案,助力企业实现降本增效闭环。综上所述,未来可期!').score)"
- ```
- ```bash
- # expected: 55
- ```
- ```bash
- node -e "const H=require('./detector/patterns.js'); console.log(H.analyzeText('今天把连接池上限从 20 调到 100,504 先压下来了。观察 24 小时,错误率 0.1% 以下就全量。').score)"
- ```
- ```bash
- # expected: 0
- ```
- ## Validation Snapshot
- Current locked examples:
- - AI-heavy zh sample: `score = 55`
- - human technical zh sample: `score = 0`
- - formal text against `casual` voice target: `voice.drift = 71`
- - current test suite: `143 CI checks`
- ## Install And Use
- ```bash
- git clone https://github.com/fendouai/humanize-text-skill.git
- cd humanize-text-skill
- npm test
- ```
- This repository is usually used as a skill package inside an agent runtime.
- - Claude Code: copy `SKILL.md`, `references/`, and `policy/` into `.claude/skills/humanize-text-skill/`
- - Cursor: use [`cursor-rules/humanize-text-skill.mdc`](./cursor-rules/humanize-text-skill.mdc)
- - Codex / OpenClaw / Hermes: point the tool at the repo root
- ### ClawHub publish readiness
- This repository is prepared for ClawHub upload:
- - canonical entry file: `SKILL.md`
- - runtime support files: `detector/`, `references/`, `policy/`
- - publish filter: [`.clawhubignore`](./.clawhubignore)
- Suggested publish command:
- ```bash
- clawhub skill publish . \
- --slug humanize-text-skill \
- --name "humanize-text-skill" \
- --version 0.1.0
- ```
- `policy/` is a required runtime dependency. If you install the skill manually
- and omit that directory, `voiceMode` will fail.
- ### Install health check
- ```bash
- node -e "const H=require('./detector/patterns.js'); const r=H.analyzeText('值得注意的是,我们打造了一套方案。', { voiceMode: 'casual' }); console.log({ score: r.score, hasVoice: !!r.voice, drift: r.voice && r.voice.drift });"
- ```
- Expected:
- - `score` is greater than `0`
- - `hasVoice` is `true`
- - `drift` is a number
- ## Common Usage Patterns
- In tools like Claude Code, Codex, or other skill-aware agents, people usually use `humanize-text-skill` in one of these ways.
- ### 1. Direct skill call
- ```text
- /humanize-text-skill Please rewrite this so it sounds less like AI:
- [paste text]
- ```
- ### 2. Natural-language request
- ```text
- Use humanize-text-skill to rewrite this so it sounds more natural:
- This project serves as a testament to our team's commitment to innovation.
- Moreover, it showcases our pivotal role in the evolving technology landscape.
- ```
- ### 3. File-based request
- ```text
- /humanize-text-skill Please humanize the copy in article.md.
- ```
- ### 4. Detect before rewrite
- ```text
- /humanize-text-skill Detect mode: show me what still sounds AI-written, but don't rewrite yet.
- ```
- ### 5. Target a voice or scene
- ```text
- /humanize-text-skill Rewrite this in a blunt voice for an issue reply.
- ```
- ```text
- /humanize-text-skill Rewrite this README intro in a technical voice.
- ```
- Practical defaults:
- - pasted text usually means `rewrite`
- - "check this first" usually means `detect`
- - "only touch this file/comment/paragraph" usually means `edit`
- - README, release note, forum post, and issue reply requests automatically benefit from scene-pack behavior
- ## Voice Calibration
- If you want the rewrite to sound more like you, provide a short sample of your own writing and use `voiceMode: custom`.
- ```text
- /humanize-text-skill
- Here's a sample of my writing for voice matching:
- [paste 2-3 paragraphs of your own writing]
- Now rewrite this in my voice:
- [paste draft]
- ```
- What the skill uses from the sample:
- - sentence length and rhythm
- - connector preferences
- - first-person tendency
- - a measurable voice profile, not private identity traits
- What it does not do:
- - it does not copy your facts or opinions from the sample
- - it does not override protected spans
- - it does not try to impersonate a real person beyond observable writing style
- ## Engine API
- ```js
- const HumanizeTextSkill = require("./detector/patterns.js");
- const result = HumanizeTextSkill.analyzeText("your text", {
- contextMode: "general",
- sceneMode: "docs",
- voiceMode: "casual",
- sample: "optional sample",
- });
- console.log(result.score);
- console.log(result.issues);
- console.log(result.voice?.drift);
- console.log(result.voice?.suggestions);
- ```
- Important:
- - `score` is the only engine score.
- - `voice.drift` is a separate dimension.
- - `fidelity` is currently enforced in the rule layer, not returned as an engine field.
- ## Skill Runtime
- 
- The detector is only one layer. The full skill runtime is:
- 1. choose mode: `rewrite`, `detect`, or `edit`
- 2. judge scene and scope
- 3. protect spans that must not drift
- 4. run detector and voice analysis
- 5. choose action: rewrite, audit-only, or minimal edit
- 6. run a second-pass residue audit
- 7. return output in a scene-appropriate shape
- See [references/skill-architecture.md](./references/skill-architecture.md) for the full map.
- ## Repository Map
- - [SKILL.md](./SKILL.md): top-level skill contract
- - [detector/](./detector/): executable detector and voice engine
- - [references/](./references/): human-readable rules, scene packs, examples, checklists
- - [policy/](./policy/): auditable scene/tier/voice policy in TOML
- - [evals/](./evals/): benchmark cases, fixtures, voice samples
- Good entry points:
- - [references/scene-packs.md](./references/scene-packs.md)
- - [references/examples.md](./references/examples.md)
- - [references/quick-checklist.md](./references/quick-checklist.md)
- - [references/voice-contract.md](./references/voice-contract.md)
- ## Related Work
- `humanize-text-skill` builds on ideas and assets from:
- - [conorbronsdon/avoid-ai-writing](https://github.com/conorbronsdon/avoid-ai-writing)
- - [MrGeDiao/shuorenhua](https://github.com/MrGeDiao/shuorenhua)
detector/core/normalize.js · Bundled script · 49 lines
- /**
- * core/normalize — bypass-trick normalization (language-agnostic).
- *
- * Migrated verbatim from avoid-ai-writing's normalizeText. Strips invisible /
- * lookalike chars that humanizer tools inject to defeat exact-string detectors,
- * BEFORE pattern matching. Tracks what was stripped so the trinary classifier
- * can use "normalization triggered" as a corroborating AI signal.
- *
- * Unicode ranges sourced from It-s-AI/llm-detection/detection/attacks/.
- */
- 'use strict';
- const CYRILLIC_LOOKALIKES = {
- 'а': 'a', 'е': 'e', 'о': 'o', 'р': 'p', 'с': 'c', 'х': 'x',
- 'у': 'y', 'к': 'k', 'м': 'm', 'н': 'h', 'в': 'b', 'т': 't',
- 'А': 'A', 'Е': 'E', 'О': 'O', 'Р': 'P', 'С': 'C', 'Х': 'X',
- 'У': 'Y', 'К': 'K', 'М': 'M', 'Н': 'H', 'В': 'B', 'Т': 'T',
- };
- const GREEK_LOOKALIKES = { 'ο': 'o', 'Ο': 'O', 'α': 'a', 'Α': 'A', 'ρ': 'p', 'Ρ': 'P' };
- const ROLEPLAY_VERBS = /^(?:nods|sighs|laughs|smiles|frowns|shrugs|grins|winks|chuckles|gasps|pauses|thinks|wonders|whispers|shouts|gestures|raises|leans|turns|looks|glances|smirks|blinks|nodding|sighing|laughing|smiling|thinking|gesturing)\b/i;
- function normalizeText(text) {
- const flags = { zeroWidth: 0, homoglyph: 0, roleplay: 0 };
- let out = text;
- // 1. Strip zero-width chars (ZWSP U+200B, ZWNJ U+200C, ZWJ U+200D,
- // BOM U+FEFF, word joiner U+2060).
- out = out.replace(/[-]/g, () => { flags.zeroWidth++; return ''; });
- // 2. Swap Cyrillic / Greek Latin-lookalike chars back to Latin.
- out = out.replace(/[Ѐ-ӿͰ-Ͽ]/g, (m) => {
- const swap = CYRILLIC_LOOKALIKES[m] ?? GREEK_LOOKALIKES[m];
- if (swap) { flags.homoglyph++; return swap; }
- return m;
- });
- // 3. Strip *roleplay-action* markers — paired *...* containing an action
- // verb anchored to the start of the inner phrase. Markdown **bold** is
- // rejected by the lookbehind/lookahead; legitimate *italic* preserved.
- out = out.replace(/(?<!\*)\*([^*\n]{1,80}?)\*(?!\*)/gu, (m, inner) => {
- if (ROLEPLAY_VERBS.test(inner)) { flags.roleplay++; return ''; }
- return m;
- });
- return { text: out, flags };
- }
- module.exports = { normalizeText };
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.