| name | ai-writing-humanizer |
| description | Remove mechanical AI flavor from English and Chinese prose while preserving facts, register, and structure. Use when a draft (blog post, essay, email, product copy, newsletter, academic prose) feels flat, robotic, or obviously AI-generated, or when asked to "humanize" / 去 AI 味 / 降低 AI 痕迹. Not for defeating AI detectors or fabricating evidence. |
AI Writing Humanizer
Remove mechanical AI flavor from English and Chinese prose. Preserve facts, register, and structure.
When to use
For drafts that feel flat, robotic, or obviously AI-generated: blog posts, essays, emails, product copy, newsletters, academic prose. Do not use to cheat detectors or fabricate evidence.
Workflow
- Diagnose — score the text against the 27 AI tells below.
- Subtract — remove mechanical patterns without adding ideas.
- Add — restore rhythm, specificity, and a human register.
Invocation
/humanize <file-or-paste> [--lang en|zh] [--tone conversational|academic|business|story|social|minimal] [--audit-only]
--lang: default auto; override with en or zh.--tone: default preserve (keep original register).--audit-only: diagnosis only, no rewrite.
Guardrails
- Do not invent facts, examples, quotes, statistics, or sources.
- Do not change meaning. Rewrite awkward but correct sentences; do not alter claims.
- Preserve register. Business stays business; academic stays academic.
- Keep structure unless it is broken or the user asks.
- Keep proper nouns, terminology, and data intact.
- When uncertain, leave it alone. False positives are better than false repairs.
Output format
## Diagnosis
- Score: X/27 (0 = human-like, 27 = heavily AI-flavored)
- Top 3 tells: ...
- Register: ...
- Risks: ...
## Subtractions
1. ...
## Additions
1. ...
## Rewritten text
...
## Final audit
- Before: X/27 | After: Y/27 | Remaining: ...
27 AI tells
Score 1 per present tell (0.5 if borderline). Full definitions and fixes are in patterns/ai-tells-en.md and patterns/ai-tells-zh.md.
Structural (1–5)
- Paragraph symmetry
- List addiction
- Transition formula
- Section bloat
- Hook-template fatigue
Lexical (6–10)
- AI buzzwords
- Hedge overload
- Empty intensifiers
- Sycophantic polish
- Faux-certainty endings
Syntactic (11–15)
- Copula avoidance
- Parallelism excess
- Tricolon addiction
- Long appositive chains
- Passive voice default
Punctuation / formatting (16–19)
- Em dash overuse
- Bold overuse
- Quote-hallucination block
- Uniform sentence length
Voice / register (20–25)
- No stance
- Reader-moralizing
- Emotion-by-label
- Generic you
- Knowledge-cutoff disclaimer
- Over-explaining
Chinese-specific (26–27)
- Translationese / 翻译腔
- Four-character slogan endings / 口号式结尾
Language rules
English
- Use concrete verbs; kill nominalizations.
- Vary sentence length; allow fragments in casual text.
- Replace "utilize", "leverage", "delve" with "use", "use", "look into".
- Drop "In this article, we will..." unless it is a real roadmap.
Chinese
- 长句先断，再调语序；拆长定语从句。
- 能用逗号不用顿号，能用句号不用分号。
- 删掉「进行」「做出」「展开」等空动词。
- 控制「的」密度；连续多个「的」必拆分。
- 口语化连接词（「其实」「不过」「说白了」）可适度使用，但要匹配文体。
- 段末不强行升华，允许开放式收尾。
Tone modes
| Mode | Direction |
|---|
conversational | shorter sentences, contractions, direct address |
academic | keep citations/precision, remove buzzwords/hedges |
business | clear action, no fluff, kill sycophancy |
story | scene, sensory detail, rhythm, point of view |
social | punchy, platform-aware (小红书 / B站 / 公众号), emoji optional, kill slogans |
minimal | only remove tells; do not inject voice |
Advanced options
--deep: second critique pass after rewrite.--voice <file.md>: adopt a writer-voice profile (e.g. 李笑来、王小波、Orwell、Didion) only if explicitly asked.--seed: benchmark the same text before and after.
Self-check
- Did I remove the top 3 detected tells?
- Did I invent any facts, examples, or quotes?
- Did I change the core meaning?
- Is the register still appropriate?
- Does it sound like the same author, just less mechanical?
- If this is Chinese, do the sentences read like a native speaker wrote them?
Fix any "no" before returning.
