Strip the AI. Keep the meaning. Sound human again.
Detects 35 AI writing patterns, scores AI probability 0-100, identifies ChatGPT / Claude / Gemini signatures, and rewrites text in your personal voice. No code. No dependencies. Pure Markdown skill for Claude Code & Claude.ai.
If this helps you write better, consider giving it a star — it helps others find it.
- Why HumanInk exists
- Features
- How HumanInk detects AI writing patterns
- Quick start / Installation
- AI probability score system
- Context modes and flags
- Before / After — Real example
- 35 AI writing patterns detected
- Style fingerprint — Write in your voice
- Multi-language AI detection
- Model signature detection
- Project structure
- FAQ
- Contributing
- References
- Changelog
LLMs write like LLMs. They inflate importance, dodge simple verbs, pad sentences with filler, and slather everything in hollow enthusiasm. Most people can feel it. They just can't name what's wrong.
HumanInk names it, scores it, reports it, then kills it.
Built on research from Wikipedia's "Signs of AI writing" and extended with 11 additional patterns from community observation. Used by writers, students, marketers, and developers who want their AI-assisted text to read like it was written by a human.
| Feature | Description |
|---|---|
| AI Score (0-100) | Quantified AI probability before and after rewriting |
| 35 AI writing patterns | Full catalog of LLM writing tells — the most comprehensive open-source list |
| Pattern report | Detailed breakdown: which patterns, how many, where |
| Context modes | --academic --casual --corporate --creative --general |
| Severity levels | --light --medium --aggressive |
| Diff output | See exactly what changed with --diff |
| Skip regions | Protect text blocks with <!-- humanink:skip --> |
| Style fingerprint | Feed your own writing samples, get output that sounds like you |
| Multi-language | AI pattern detection in 7 languages with calibrated scoring |
| Language shortcuts | --pt and --es to force language mode directly |
| Explain mode | --explain adds inline annotations showing why each change was made |
| Claude.ai compatible | Works as Claude Code skill or Claude.ai Project skill |
| Context-aware scoring | Severity adjusts per mode — academic text isn't penalized for "Additionally" |
| Model signatures | Identifies ChatGPT vs Claude vs Gemini pattern clusters |
| Paragraph scoring | Per-paragraph breakdown shows where AI crept in |
| Confidence indicator | High/Medium/Low confidence based on unique pattern count |
| Rewrite guardrails | Prevents over-correction — preserves facts, word count, intent |
┌──────────────┐
Your text ──────────>│ Parse flags │
└──────┬───────┘
│
┌──────v───────┐
│ AI Score │──────> Score: 73/100 (Likely AI)
│ (0-100) │
└──────┬───────┘
│
┌──────v───────┐
│ 35-pattern │──────> Pattern report
│ scan │ (#7 AI vocab ×6, #1 inflation ×3...)
└──────┬───────┘
│
┌──────v───────┐
│ Rewrite │──────> Draft (mode + severity applied)
│ + voice │
└──────┬───────┘
│
┌──────v───────┐
│ Anti-AI │──────> "What still screams AI?"
│ audit │
└──────┬───────┘
│
┌──────v───────┐
│ Re-score │──────> Score: 73 → 8 (-65)
└──────┬───────┘
│
┌──────v───────┐
│ Deliver │──────> Final version + report + diff
└──────────────┘
Install (one command):
mkdir -p ~/.claude/skills && git clone https://github.com/sirambrosio/humanink.git ~/.claude/skills/humaninkUse in Claude Code:
/humanink
[paste your text]
With flags:
/humanink --casual --aggressive --diff --report
[paste your text]
Update to latest version:
cd ~/.claude/skills/humanink && git pullClaude.ai (Projects) installation
- Download or ZIP the
humaninkfolder - Go to Claude.ai → Projects → Project Knowledge
- Upload the files (
SKILL.md,PATTERNS.md,STYLEGUIDE.md) - Use
/humaninkin chat
Every text gets a score from 0 to 100 before and after rewriting. The algorithm uses pattern count, severity weights, density multiplier, context-aware adjustment, overlap correction, and a confidence indicator.
╔══════════════════════════════════╗
║ AI SCORE: 73 → 8 (-65) ║
║ Status: Human ║
╚══════════════════════════════════╝
| Score | Label | Meaning |
|---|---|---|
| 0-15 | Human | No significant AI patterns detected |
| 16-35 | Mostly human | Minor traces, could be coincidental |
| 36-55 | Mixed | Noticeable AI patterns, needs editing |
| 56-75 | Likely AI | Clear AI signatures throughout the text |
| 76-100 | AI slop | Unmistakably machine-generated content |
Paragraph-level scoring shows which sections are AI-written vs human-written — useful for detecting mixed-authorship documents. See PATTERNS.md for pattern details and SKILL.md for the full scoring algorithm with a worked example.
Choose one to match your writing context. Each mode adjusts pattern severity — for example, academic mode won't penalize "Additionally" since it's standard in papers.
| Flag | Tone | Best for |
|---|---|---|
--general |
Balanced, natural (default) | Most text |
--academic |
Formal, precise, citations preserved | Papers, research, essays |
--casual |
Loose, contractions, first person | Blog posts, social media |
--corporate |
Clean, confident, professional | Reports, decks, emails |
--creative |
Maximum voice, edge, rule-breaking | Essays, opinion pieces |
| Flag | What it does |
|---|---|
--light |
Fix only the worst offenders (severity 8-10) |
--medium |
Fix all clear AI patterns (default) |
--aggressive |
Full rewrite, restructure, maximize human voice |
| Flag | What it adds |
|---|---|
--diff |
Shows removed/added text in diff format |
--report |
Full pattern report with counts and locations |
--score |
Score only, skip the rewrite |
--no-audit |
Skip second-pass audit (faster) |
--explain |
Inline annotations explaining why each change was made |
--pt |
Force Portuguese mode (skip auto-detection) |
--es |
Force Spanish mode (skip auto-detection) |
| Feature | How to use |
|---|---|
| Skip regions | Wrap text in <!-- humanink:skip --> ... <!-- humanink:end --> |
| Style fingerprint | Use --style then paste your own writing samples first |
|
INPUT |
OUTPUT |
|---|---|
|
|
| 15 pattern instances found across 11 patterns | 0 patterns remaining · AI Score dropped 72 points |
The most comprehensive open-source catalog of AI writing tells. Each pattern includes detection triggers, severity weights, before/after examples, and cross-language equivalents.
Content patterns #1-6
| # | Pattern | Sev | AI tells |
|---|---|---|---|
| 1 | Significance inflation | 9 | "marking a pivotal moment in the evolution of..." |
| 2 | Notability name-dropping | 7 | "cited in NYT, BBC, FT, and The Hindu" |
| 3 | Superficial -ing analyses | 8 | "symbolizing... reflecting... showcasing..." |
| 4 | Promotional language | 8 | "nestled within the breathtaking region" |
| 5 | Vague attributions | 7 | "Experts believe it plays a crucial role" |
| 6 | Formulaic challenges | 6 | "Despite challenges... continues to thrive" |
Language patterns #7-12
| # | Pattern | Sev | AI tells |
|---|---|---|---|
| 7 | AI vocabulary | 9 | "Additionally... testament... landscape..." |
| 8 | Copula avoidance | 7 | "serves as... features... boasts" |
| 9 | Negative parallelisms | 6 | "It's not just X, it's Y" |
| 10 | Rule of three | 5 | "innovation, inspiration, and insights" |
| 11 | Synonym cycling | 6 | "protagonist... main character... hero..." |
| 12 | False ranges | 5 | "from the Big Bang to dark matter" |
Style patterns #13-18
| # | Pattern | Sev | AI tells |
|---|---|---|---|
| 13 | Em dash abuse | 5 | "this — and that — yet also —" |
| 14 | Bold overuse | 4 | "OKRs, KPIs, BMC" |
| 15 | Inline-header lists | 6 | "Performance: Performance improved" |
| 16 | Title Case headings | 4 | "Strategic Negotiations And Partnerships" |
| 17 | Emojis in pro text | 5 | ":rocket: Launch: :bulb: Insight:" |
| 18 | Curly quotes | 3 | \u201clike this\u201d |
Communication & filler patterns #19-24
| # | Pattern | Sev | AI tells |
|---|---|---|---|
| 19 | Chatbot artifacts | 10 | "I hope this helps! Let me know..." |
| 20 | Cutoff disclaimers | 8 | "While details are limited..." |
| 21 | Sycophantic tone | 9 | "Great question! Absolutely!" |
| 22 | Filler phrases | 6 | "In order to", "Due to the fact that" |
| 23 | Excessive hedging | 7 | "could potentially possibly" |
| 24 | Generic conclusions | 8 | "The future looks bright" |
Extended patterns #25-35
| # | Pattern | Sev | AI tells |
|---|---|---|---|
| 25 | "In today's world" openers | 8 | "In today's rapidly evolving landscape..." |
| 26 | Dead metaphors | 6 | "double-edged sword", "tip of the iceberg" |
| 27 | "Meanwhile" transitions | 5 | Overuse of "Meanwhile", "That said", "On the other hand" |
| 28 | Mirror conclusions | 7 | Conclusion paraphrases the introduction |
| 29 | Uniform paragraph length | 6 | Every paragraph is exactly 3 sentences |
| 30 | "It is worth noting" hedges | 7 | "It should be noted that...", "It bears mentioning..." |
| 31 | Exhaustive list syndrome | 6 | 8-item lists where 3 items would do |
| 32 | Rubber stamp qualifiers | 5 | "significant", "notable", "remarkable" as filler |
| 33 | "Let's dive in" opener | 8 | "Let's dive in", "Let's explore", "Without further ado" |
| 34 | Bullet-heavy formatting | 5 | >50% content in bullets where prose fits better |
| 35 | Unnecessary numbered steps | 5 | Numbered steps for things that aren't procedures |
Full pattern documentation with detection triggers, before/after examples, cross-language equivalents, and scoring weights: PATTERNS.md
Feed HumanInk samples of your own writing. It analyzes your sentence length, vocabulary, punctuation habits, paragraph rhythm, and tone — then rewrites AI text in your voice.
/humanink --style
[paste 2-3 samples of your writing, minimum 200 words]
[then paste the text you want rewritten]
╔══════════════════════════════════════╗
║ VOICE PROFILE ║
╠══════════════════════════════════════╣
║ Sentences: short-medium (avg 14w) ║
║ Variance: high (SD 6.2) ║
║ Contractions: always ║
║ Vocabulary: mid-level, some jargon ║
║ Humor: dry, infrequent ║
║ Directness: high ║
╚══════════════════════════════════════╝
Full voice analysis framework: STYLEGUIDE.md
The 35 AI writing patterns are universal across languages. HumanInk detects language automatically and maps patterns to local equivalents. Use --pt or --es to skip detection and force a language.
Supported languages: English, Portuguese, Spanish, French, German, Japanese, Italian.
See language-specific AI tells
| Language | Common AI tells |
|---|---|
| 🇧🇷 Portuguese | "Adicionalmente", "destaca-se", "fomentar", "paisagem" (abstract), "No mundo atual" |
| 🇪🇸 Spanish | "Adicionalmente", "panorama", "destacar", "fomentar", "En el mundo actual" |
| 🇫🇷 French | "De plus", "mettre en lumiere", "au coeur de", "Dans le monde d'aujourd'hui" |
| 🇩🇪 German | "Daruber hinaus", "entscheidend", "Landschaft" (abstract), stacked subordinate clauses |
| 🇯🇵 Japanese | Overuse of "において", excessive "という" nominalizations, "さらに" |
| 🇮🇹 Italian | "Inoltre", "sottolineare", "panorama", "nell'ambito di" |
Scoring thresholds adapt per language — em dashes score lower in Portuguese (where they're common), formal register scores differently in academic Japanese, etc.
HumanInk identifies which AI model likely generated the text based on pattern clusters:
| Model | Signature patterns | Common tells |
|---|---|---|
| ChatGPT | Formatting-heavy (#14, #15, #17, #34) | Bold overuse, emoji lists, bullet-heavy output |
| Claude | Hedging-heavy (#22, #23, #30) | Excessive qualifiers, "It's worth noting", over-cautious tone |
| Gemini | Inflation-heavy (#1, #4, #25) | "In today's world", promotional language, significance inflation |
Model detection is reported as a hint, not a definitive claim — patterns overlap between models.
humanink/
├── SKILL.md ← core skill definition (source of truth)
├── PATTERNS.md ← 35 patterns: triggers, examples, cross-language, scoring
├── STYLEGUIDE.md ← voice fingerprint framework
├── WARP.md ← guidance for Warp terminal
└── README.md ← you are here
No runtime. No dependencies. No build step. No API keys. Pure Markdown. Works as a Claude Code skill or Claude.ai Project skill.
Does HumanInk work with ChatGPT or other LLMs?
HumanInk is designed as a skill for Claude Code and Claude.ai. It doesn't work with ChatGPT directly. However, the pattern catalog in PATTERNS.md is model-agnostic — you can use it as a reference to manually identify AI patterns in any text.
Will this bypass AI detectors like GPTZero, Turnitin, or Originality.ai?
HumanInk rewrites text to remove recognizable AI patterns and match human writing characteristics. The result reads more naturally because it is more natural — not because it's "tricking" a detector. The focus is on writing quality, not detection evasion.
Can I use this for academic work?
Use --academic mode. It preserves citations, formal register, and field-specific terminology while removing the robotic patterns that make AI-assisted text obvious. Always follow your institution's AI usage policies.
How is this different from Undetectable AI, QuillBot, etc.?
Those are paid SaaS tools that paraphrase text. HumanInk is different:
- Open source and free forever
- Transparent — you can read every pattern and scoring rule
- No API — runs entirely inside Claude as a skill
- Style-aware — can replicate your voice, not just generic "human" text
- Educational —
--explainmode teaches you why each pattern is AI-like
Does it work in languages other than English?
Yes. HumanInk supports 7 languages: English, Portuguese, Spanish, French, German, Japanese, and Italian. Each language has calibrated scoring thresholds and localized AI pattern equivalents.
Contributions are welcome. Here's how you can help:
- Report a new AI pattern — found a pattern not in the list? Open an issue with examples
- Add language support — extend cross-language equivalents in PATTERNS.md
- Improve detection — suggest better triggers or before/after examples
- Fix bugs — found a scoring edge case? Submit a PR
See WARP.md for the full guide on adding patterns and editing files.
- Wikipedia: Signs of AI writing — primary source for patterns 1-24
- WikiProject AI Cleanup — the team maintaining the guide
- Patterns 25-35 sourced from community observation and extended research
"LLMs use statistical algorithms to guess what should come next. The result tends toward the most statistically likely result that applies to the widest variety of cases." — Wikipedia
| Version | What changed |
|---|---|
| 2.2.0 | Context-aware severity (per mode), model signatures (ChatGPT/Claude/Gemini), paragraph-level scoring, confidence indicator, rewrite guardrails, co-occurrence detection, Italian language support, worked scoring example |
| 2.1.0 | 35 patterns (#33-#35 new), --pt/--es language shortcuts, --explain mode, Claude.ai compatibility, overlap scoring groups, threshold gating |
| 2.0.0 | AI score system, 32 patterns (8 new), pattern report, context modes, severity levels, diff output, skip regions, style fingerprint, multi-language calibration |
| 1.0.0 | Initial release. 24 patterns, two-pass audit. |
HumanInk
by Andre Ambrosio
If a machine wrote it, a machine can un-write it.
