# humanizer-semitic
AI text humanizer skills for Arabic and Hebrew — built for Claude Code, Cursor, Windsurf, and every agent that supports the [skills](https://skills.sh) protocol.
Four independent skills. One repo. Zero AI tells.
---
## Why this exists
LLMs write Arabic and Hebrew the same way they write everything else — with the wrong words, the wrong rhythm, and none of the cultural instincts that make text feel like it came from a human being.
For Semitic languages, the failure is worse than for European languages. Arabic has multiple registers and dialects with wildly different grammars. Hebrew encodes gender into every verb and adjective. Modern Standard Arabic AI text reads like a government report. Egyptian AI text sounds like MSA wearing a costume. Levantine AI text scores **1.3 BLEU** on generation benchmarks — a 17x gap from the reverse direction. Hebrew AI text gets the gender wrong and calls it done.
These skills fix that. Each one targets a specific language or dialect with patterns grounded in academic research, native speaker feedback, and stylometric analysis.
---
## Skills
| Skill | Language | Patterns | Install |
|---|---|---|---|
| `humanizer-ar-msa` | Modern Standard Arabic (الفصحى) | 28 | `npx skills add OthmanAdi/humanizer-semitic@humanizer-ar-msa` |
| `humanizer-ar-egt` | Egyptian Arabic (عامية مصرية) | 25 | `npx skills add OthmanAdi/humanizer-semitic@humanizer-ar-egt` |
| `humanizer-ar-shami` | Levantine Arabic — Syrian, Lebanese, Palestinian (الشامي) | 25 | `npx skills add OthmanAdi/humanizer-semitic@humanizer-ar-shami` |
| `humanizer-he` | Modern Hebrew (עברית מודרנית) | 35 | `npx skills add OthmanAdi/humanizer-semitic@humanizer-he` |
---
## Installation
Install a single skill:
```bash
npx skills add OthmanAdi/humanizer-semitic@humanizer-ar-msa
npx skills add OthmanAdi/humanizer-semitic@humanizer-ar-egt
npx skills add OthmanAdi/humanizer-semitic@humanizer-ar-shami
npx skills add OthmanAdi/humanizer-semitic@humanizer-he
```
Browse and pick interactively:
```bash
npx skills add OthmanAdi/humanizer-semitic
```
Or from npm, all four skills in one package ([humanizer-semitic](https://www.npmjs.com/package/humanizer-semitic)):
```bash
pi install npm:humanizer-semitic
```
Install everything:
```bash
npx skills add OthmanAdi/humanizer-semitic --all
```
Install globally (available across all projects):
```bash
npx skills add OthmanAdi/humanizer-semitic@humanizer-ar-egt -g
```
---
## Usage
Once installed, trigger a skill by mentioning it in your agent:
- `/humanizer-ar-msa` — humanize this MSA text
- `/humanizer-ar-egt` — humanize this Egyptian Arabic text
- `/humanizer-ar-shami` — humanize this Levantine Arabic text
- `/humanizer-he` — humanize this Hebrew text
Or naturally: *"make this sound like a human wrote it in Egyptian Arabic"*
---
## What each skill does
Each skill runs a three-stage workflow:
1. **Identify** — scans for AI-generated patterns specific to the language or dialect
2. **Rewrite** — fixes each pattern while preserving meaning and appropriate register
3. **Audit** — final pass to catch remaining AI tells and calibrate voice
Every skill includes a 50-point quality rubric so the agent can self-assess the output.
---
## Skill highlights
### `humanizer-ar-msa` — Modern Standard Arabic
Targets 28 patterns across five categories: hedging language, lexical uniformity, syntactic templates, rhetorical deficiencies, and diacritics. Specific focus on the grammatically incorrect `علاوة على ذلك` transition (which AI uses constantly despite being wrong), the absence of `السجع` (rhymed prose), and the `تم/يتم` passive voice overload.
> AI uses syntactic templates 95% of the time. Human Arabic writers: 38%.
### `humanizer-ar-egt` — Egyptian Arabic
Targets 25 patterns. Egyptian is the only Arabic dialect understood across all 22 Arab countries — making AI failures here affect over 400 million speakers. Key fixes: MSA vocabulary substitution table (الآن → دلوقتي, أريد → عايز), demonstrative order flip (هذا الكتاب → الكتاب ده), circumfix negation restoration (ما...ش), and discourse particle injection (يعني، بقى، خلاص).
> AI almost never produces the Egyptian circumfix negation pattern `ما...ش`. Its absence is one of the strongest AI tells in colloquial Arabic.
### `humanizer-ar-shami` — Levantine Arabic
Targets 25 patterns with regional calibration for Syrian, Lebanese, and Palestinian variants. Covers the full verb morphology gap (ب-prefix, عم progressive, رح future, بدّ for "want"), code-switching (French for Lebanese, English for Syrian/Palestinian), and dialect-specific orthography.
> Research benchmark: LLMs score 1.3 BLEU generating Levantine Arabic — versus 23 BLEU on the reverse direction.
### `humanizer-he` — Modern Hebrew
Targets 35 patterns across five categories. Heavy focus on grammatical gender agreement (AI defaults to masculine), construct state (סמיכות) misuse, nikud (vowel points) overuse, register collapse, and the anglicization of word order.
---
## Related skills
These skills are inspired by and structurally follow [blader/humanizer](https://github.com/blader/humanizer) (12.9k stars) — the original English AI text humanizer.
Other language variants:
- [op7418/Humanizer-zh](https://github.com/op7418/Humanizer-zh) — Chinese Simplified
- [kevintsai1202/Humanizer-zh-TW](https://github.com/kevintsai1202/Humanizer-zh-TW) — Chinese Traditional
- [gonta223/humanizer-ja](https://github.com/gonta223/humanizer-ja) — Japanese
- [DaleSeo/korean-skills](https://github.com/DaleSeo/korean-skills) — Korean
- [marmbiz/humanizer-de](https://github.com/marmbiz/humanizer-de) — German
---
## License
MIT