arabic-nlp

name	arabic-nlp
description	Use this skill when selecting or applying Arabic NLP, tokenization, stemming, diacritization, morphological analysis, stopword, OCR, speech, or language-model tools including CAMeL Tools, Farasa, PyArabic, Tashaphyne, Qalsadi, and Arabic corpora.

Arabic NLP

When To Use

Use this skill for arabic-nlp work in Arab/MENA contexts, especially when the prompt includes: Arabic NLP, tokenization, stemming, diacritization, morphology, Farasa, CAMeL, PyArabic.

When Not To Use

• Do not claim model accuracy without benchmark sources.
• Do not send private text to third-party APIs without approval.

Required Inputs

• Country or market.
• Target vendor, if already selected.
• Desired workflow: vendor selection, implementation, review/debugging, launch readiness, or source research.
• Sandbox vs production state.
• Whether live customer, payment, tax, identity, bank, or payroll data is involved.

Common Workflows

• Identify the NLP task and deployment constraints.
• Prefer maintained libraries and official docs or GitHub repos.
• Separate API services from local libraries.
• Return evaluation data needs and dialect/classical Arabic caveats.

Default Workflow

1. Read sources.yml to see available vendors and confidence.
2. If a vendor is named, read only vendors/<vendor-id>.md for that vendor.
3. If choosing vendors, compare only vendors in this skill's registry: arabic-stopwords, camel-tools, farasa, pyarabic, qalsadi, tashaphyne.
4. Load references/integration-checklist.md only for implementation, review, or launch-readiness work.
5. Use scripts/list-vendors.mjs for a deterministic vendor list when needed.
6. Answer with source-backed facts, explicit unknowns, and validation steps.

Decision Tree

• Named vendor: read that vendor file, then answer narrowly.
• Vendor selection: filter by country, docs access, maturity, and source quality before recommending.
• Implementation: include auth, sandbox, webhook/callback, retries, idempotency, logging, and error handling only where source-backed.
• Review/debugging: compare the user's plan or code against the vendor file, sources.yml, and references/integration-checklist.md.
• Source research: update facts only when an official source, developer portal, GitHub repo, OpenAPI/Postman asset, or government source supports the claim.
• Missing docs: say Needs vendor access or Unknown from public docs.

Response Contract

• Start by naming the skill file and vendor/reference files used.
• Give a short recommendation or implementation path before details.
• Separate source-backed facts from assumptions and unknowns.
• Include country/market fit, docs access, docs confidence, and source-quality caveats when selecting vendors.
• Include a validation checklist with sandbox/test steps, rollback or retry notes, and manual approval gates for high-risk work.
• Never provide live-action instructions that move money, tax documents, bank data, identity data, payroll data, or outbound messages without explicit human approval.

Files To Read

• Routing and process: SKILL.md.
• Vendor facts: vendors/*.md.
• Source map: sources.yml.
• Implementation review: references/integration-checklist.md.
• Example response style: examples/source-backed-answer.md.

Safety Rules

• Check license and data handling before production use.
• Avoid sending sensitive Arabic text to external APIs by default.
• Benchmark on the target dialect and domain.
• Document tokenizer/stemmer limitations.

Validation Checklist

• Vendor facts map back to source_urls.
• Unknowns are labeled instead of guessed.
• Sandbox and production are separated.
• Secrets are not printed or committed.
• High-risk live actions require explicit human approval.
• Evals in evals/prompts.yml still cover the changed workflow.

Done Criteria

• The answer names the files read or source-backed references used.
• The implementation plan includes tests and rollback/verification steps.
• No unsupported regional, API, compliance, pricing, or endpoint claims are included.