Skip to content

baldiga/hebrew-writer

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

32 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

הכותב העברי

Hebrew Writer — Claude Code Skill

Write Hebrew that passes every AI detector and fools native Israelis.

License: MIT Claude Code Lines Hebrew


The first and most comprehensive Hebrew-native writing skill for Claude Code. Built on analysis of 50+ repos, 10+ academic papers, 6 detection tools, and 500K real Israeli comments.


What This Does

You give it a topic. It writes Hebrew that sounds like a native Israeli sat down and typed it.

Not translated English. Not formal textbook Hebrew. Not "AI Hebrew" that every Israeli spots in two seconds. Real, casual, dugri Israeli Hebrew — the kind you'd read on Geektime, in a LinkedIn post from a Tel Aviv founder, or in a WhatsApp message from your team lead.

It passes GPTZero. It passes Originality.ai. It passes ZeroGPT, Copyleaks, and Turnitin. And more importantly — it passes the test that actually matters: a native Israeli reading it and not thinking twice.

How It Works — 7 Layers

Layer What It Does
Hebrew Mind Forces Claude to think and compose in Hebrew from scratch — not translate from English
Anti-Detection Engine 55+ AI patterns mapped to Hebrew, including 16-word blacklist, 6 content patterns, and 10 Hebrew-specific tells no English humanizer covers
Israeli Voice Dugri directness, slang system (Arabic/Yiddish/English loanwords), discourse markers (כאילו, יעני, תכל'ס), humor, cultural texture
Linguistic Precision Gender system with strategic imperfection, spelling variation, construct state vs. של, binyanim awareness, pro-drop
Rhythm Engineering Data-driven sentence lengths (avg 8.9 words, based on 500K real comments), burstiness, perplexity injection, ellipsis, n-gram breaking
Self-Audit Loop 8-dimension scoring (95/100 threshold), autonomous quality gate, no user intervention needed
Voice Matching Clone any person's Hebrew writing voice from samples, save persistent profiles

The Big No-Nos

Five advanced AI patterns that survive every other humanizer — we catch them all:

  1. Macro Feeling Copy — "ויש לזה מחיר אמיתי:" is a drum roll, not content. We delete it.
  2. Presentation Slide Structure — Parallel stacked points pretending to be prose. We weave them into the argument.
  3. LinkedIn Punchline Syndrome — "הפתרון הכי חכם הוא הפתרון הכי פשוט" belongs on a sunset poster, not in writing.
  4. Disconnected Temperature — Flat emotional energy throughout. Real writers care more about some parts.
  5. "Not X, It's Y" Addiction — The most common AI essay structure. We break it.

Quick Start

Install (Global — works everywhere)

mkdir -p ~/.claude/skills/hebrew-writer
curl -sL https://raw.githubusercontent.com/baldiga/hebrew-writer/main/SKILL.md \
  -o ~/.claude/skills/hebrew-writer/SKILL.md
curl -sL https://raw.githubusercontent.com/baldiga/hebrew-writer/main/voice-profile-template.md \
  -o ~/.claude/skills/hebrew-writer/voice-profile-template.md

Install (Project-scoped)

mkdir -p .claude/skills/hebrew-writer
curl -sL https://raw.githubusercontent.com/baldiga/hebrew-writer/main/SKILL.md \
  -o .claude/skills/hebrew-writer/SKILL.md
curl -sL https://raw.githubusercontent.com/baldiga/hebrew-writer/main/voice-profile-template.md \
  -o .claude/skills/hebrew-writer/voice-profile-template.md

Use It

# Generate from scratch
/hebrew-writer "למה סטארטאפים ישראלים מצליחים"

# Rewrite AI-sounding Hebrew
/hebrew-writer "הטכנולוגיה המתקדמת מהווה גורם מכריע..." --mode rewrite

# Scan text for AI tells
/hebrew-writer "paste your text here" --mode detect

# Clone someone's voice
/hebrew-writer --learn "paste their writing" --save-as avi
/hebrew-writer "topic" --voice avi

# Control length, type, gender
/hebrew-writer "topic" --length long --type academic --gender female

Modes

Mode What It Does
generate Write from scratch. Give it a topic, get a complete piece.
rewrite Transform existing content (Hebrew, English, or mixed) into natural Israeli Hebrew.
detect Scan text and report AI patterns found. No changes — diagnosis only.

Content Types

Auto-detected or manually set with --type:

Type Register Example
blog Casual Israeli (3/10 formality) Blog posts, LinkedIn, articles
academic Formal with Israeli backbone (8/10) Papers, research, academic writing
social Maximum casual (1/10) Twitter, Facebook, Instagram
business Professional but Israeli (6/10) Reports, proposals
email Direct and terse (4/10) Israeli business emails
creative Adapts to story voice Fiction, poetry, narrative

Voice Cloning

Feed it a sample of someone's writing. It learns their patterns and writes as them.

# Learn from inline text (minimum 200 words)
/hebrew-writer --learn "paste 200+ words of their writing here" --save-as danny

# Learn from a file
/hebrew-writer --learn --my-voice-file path/to/writing.md --save-as danny

# Learn from multiple files (highest accuracy)
/hebrew-writer --learn --my-voice-files path/to/folder/ --save-as danny

# Use saved voice
/hebrew-writer "topic" --voice danny

Accuracy tiers: 200-500 words = basic match | 500-1500 = strong | 1500+ = full clone | multiple files = maximum

Self-Audit Scoring

Every piece is internally scored on 8 dimensions before output:

Dimension Weight What It Measures
דוגריות (Directness) 13% Does it take a position?
קצב (Rhythm) 14% Sentence length variation
אמינות (Authenticity) 14% Could an Israeli have written this?
טקסטורה (Texture) 11% Slang, register shifts, word surprises
נשמה (Soul) 12% Emotions, opinions, visible thinking
צפיפות (Density) 8% Every sentence earns its place
רישום (Register) 8% Formality matches content type
אנטי-זיהוי (Anti-Detection) 20% Statistical fingerprint clean

Threshold: 95/100. No dimension below 8/10. Fully autonomous — no mid-process prompts.

Add --show-score to see the breakdown.

Research Foundation

This skill wasn't built on vibes. It's backed by:

  • 6 AI detection tools analyzed — GPTZero, Originality.ai, Turnitin, Copyleaks, ZeroGPT, Sapling
  • 10+ academic papers — DetectGPT (ICML 2023), Binoculars (ICML 2024), RAID Benchmark (ACL 2024), and more
  • 50+ GitHub repos reviewed — blader/humanizer, humanizalo, stop-slop, anti-ai-slop-writing, unslop
  • 500,000 real Israeli comments analyzed — HeBERT UGC dataset from Tel Aviv University
  • Real Geektime articles analyzed — for authentic Israeli tech writing patterns
  • Hebrew linguistics research — Academy of Hebrew Language standards, discourse marker studies, morphological analysis

What Makes This Different

Feature Other Humanizers Hebrew Writer
Language English only Hebrew-native from the ground up
Hebrew-specific tells None 10 patterns no English skill covers
Data-driven Based on theory Based on 500K real Israeli comments
Cultural authenticity N/A Israeli dugri culture, slang, humor
Sentence length Theory-based (15-20 words) Data-driven (avg 8.9 words)
Voice cloning Some Full system with persistent profiles
Quality threshold 42/60 or 80/100 95/100 with 8-dimension scoring
Advanced patterns Vocabulary + formatting Big No-Nos: macro copy, slide structure, punchlines, temperature, X-vs-Y

Compatibility

Works with Claude Code, Cursor, VS Code, Codex CLI, Gemini CLI, Windsurf, Continue.dev, and OpenClaw.

Files

File Lines What It Is
SKILL.md 2,081 The complete skill — all 7 layers
voice-profile-template.md 118 Template for saved voice profiles

License

MIT — use it, share it, improve it.


Built with obsessive attention to what makes Hebrew sound human.

About

The first Hebrew-native writing skill for Claude Code. Passes all AI detectors. Fools native Israelis. 7-layer system with 55+ pattern detection, Israeli voice injection, and voice cloning.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors