claude-arabic-writing

claude-arabic-writing

A skill for Claude that makes it write and translate Arabic that reads correctly and naturally — grammatically sound, free of typos, idiomatic, and free of the tell-tale signs of machine translation.

It covers Modern Standard Arabic (الفصحى) plus four major dialects — Egyptian, Levantine, Gulf, and Maghrebi — distilling authoritative grammar, orthography, style, and localization references into guidance Claude applies automatically while it writes.

Why you might want it

Out of the box, LLM Arabic often looks fluent but reads as translated: من قبل passives, قام بـ + verbal noun, الخاص بك for every "your", drifting gender/number agreement, calqued collocations, and Latin punctuation. This skill encodes the rules native editors apply, so the Arabic Claude produces for you sounds like a person wrote it — not a machine.

With the skill installed, Claude will:

• Write original Arabic prose in the right register for your audience.
• Translate into Arabic (especially from English) by re-composing meaning, not transferring English word-for-word.
• Edit / proofread Arabic for hamza, ة/ه, ى/ي, agreement, iḍāfa, numbers, punctuation, collocations, and naturalness.
• Pick the right variety — formal MSA for documents, or a specific dialect for marketing and conversational copy — without mixing them.

See it in action — أمثلة

Arabic below renders right-to-left. The ❌ column is typical machine/LLM output; the ✅ column is what the skill steers Claude toward.

Removing the "machine-translated" tells

Each ✅ fixes a specific tell: active voice instead of من قبل, a single verb instead of قام بـ, an attached suffix instead of الخاص بـ, feminine-singular agreement for a non-human plural, the internal passive instead of يتم, and a fronted verb instead of the copular calque.

Translating a full sentence (English → natural MSA)

English: We design, build, and operate the systems that mid-sized businesses rely on — from cloud platforms to autonomous AI.

العربية: نُصمِّم الأنظمةَ التي تعتمد عليها الشركات المتوسطة ونبنيها ونُشغّلها — من المنصّات السحابية إلى الذكاء الاصطناعي المستقلّ.

Note the verb-first sentence, the relative clause with a resumptive pronoun (التي تعتمد **عليها**), and the non-human plural الأنظمة taking التي.

The same message in MSA and each dialect

English: I want to go now — where's the restaurant?

Notice how the verb, the word for "want", "now", and "where" all change per dialect — and that Egyptian moves the question word to the end.

Proofreading

Input: انا ذهبت الى المدرسه , و قابلت ثلاث طلاب هناك

Corrected: أنا ذهبت إلى المدرسة، وقابلت ثلاثة طلاب هناك.

Fixes: انا → أنا and الى → إلى (hamza), المدرسه → المدرسة (ة), ثلاث → ثلاثة (number polarity — a masculine counted noun takes the ة), plus the Arabic comma ، with no space before it and the attached و.

Requirements

• Claude Code (or any client that supports Agent Skills via a SKILL.md).
• git (to clone) — or you can download the repo as a ZIP.

Installation

For all your projects (personal)

git clone https://github.com/ahmeddabak/claude-arabic-writing.git
cp -r claude-arabic-writing/skills/arabic-writing ~/.claude/skills/arabic-writing

For a single project

git clone https://github.com/ahmeddabak/claude-arabic-writing.git
mkdir -p .claude/skills
cp -r claude-arabic-writing/skills/arabic-writing .claude/skills/arabic-writing

Start a new Claude Code session afterward so the skill is picked up. That's it — no build step, no dependencies.

Usage

The skill activates automatically whenever you ask for Arabic. You don't need to mention it by name. For example:

• "Translate this product description into natural MSA."
• "Rewrite this welcome message in Gulf Arabic for a Saudi audience."
• "Proofread this Arabic paragraph and list the fixes."
• "Say this in Egyptian Arabic: …"
• "This Arabic reads like a machine translated it — make it sound native."

You can also invoke it explicitly with /arabic-writing.

What's inside

skills/arabic-writing/
├── SKILL.md                 # entry point: workflow + correctness checklist + "AI smell test"
├── references/
│   ├── grammar.md           # word order, إعراب, agreement, iḍāfa, numbers
│   ├── orthography.md       # hamza, ة/ه, ى/ي, tanwin, punctuation, diacritics, typos
│   ├── translationese.md    # machine-translation tells + English→Arabic pitfalls
│   ├── collocations.md      # awkward-calque → natural tables + verb–preposition government
│   ├── register-and-style.md# register choice, connectors, naturalness, tone
│   ├── conventions.md       # transliteration, numerals, dates, units, bidi
│   └── dialects/
│       ├── egyptian.md
│       ├── levantine.md
│       ├── gulf.md
│       └── maghrebi.md
└── evals/
    └── evals.json           # objective test cases

It's built on progressive disclosure: SKILL.md stays short (a workflow, a top-10 correctness checklist, and a "smell test" for machine-made Arabic), and the detailed rules live in references/ files that Claude reads only when relevant — so the skill is thorough without bloating Claude's context.

Testing it yourself

skills/arabic-writing/evals/evals.json contains objective test cases — each with a prompt and checkable expected behaviors — covering agreement, the قام بـ / من قبل tells, possessives, proofreading, register choice, dialect fidelity, and numbers/dates. Run any of them with and without the skill to see the difference.

Caveats

• The bundled Arabic examples illustrate documented patterns; some pairs are constructed rather than quoted. A native-speaker review is recommended before relying on the output for published material.
• A few orthographic choices have more than one accepted convention (مسؤول/مسئول، مئة/مائة).
• For high-stakes localization, target a specific country (Egyptian, Saudi, Moroccan, …) rather than "Arabic" in general.

Contributing

Issues and pull requests are welcome — corrections from native speakers and professional translators especially. Good contributions include: additional calque→natural collocation pairs, dialect refinements, and new eval cases that expose a failure mode.

License

MIT — free to use, modify, and redistribute.