A2 · AGENT SKILLS / مهارات الوكلاء

Build one lean, reusable Skill.

Turn a repeated bilingual review workflow into a discoverable SKILL.md with supporting detail loaded only when needed. حوّل سير مراجعة متكرر إلى مهارة موجزة قابلة لإعادة الاستخدام.

AI for Palestine practice template—not an official Anthropic resource.
قالب تدريبي من AI for Palestine وليس مورداً رسمياً من Anthropic.

01

Install, inspect, then test the trigger.

  1. 01

    Create the folder

    Place both downloads in .claude/skills/bilingual-ui-review/. Commit project skills only after reviewing every instruction and tool grant.

  2. 02

    Test a matching request

    Ask: “Review this English/Arabic interface for meaning, direction and overflow.” Confirm the Skill loads and returns file-and-line evidence.

  3. 03

    Test a non-matching request

    Ask an unrelated coding question. The Skill should not activate just because the repository contains Arabic text.

  4. 04

    Keep progressive disclosure

    Keep the entry file short. Move deep rules, examples and deterministic helpers into supporting files and reference them explicitly.

02

Download the starter files.

SKILL.md

Download SKILL.md ↓
---
description: Reviews student-facing English and Arabic UI copy...
allowed-tools: Read, Grep, Glob
---

# Bilingual UI review
1. Read checklist.md before reviewing.
2. Compare meaning, direction and overflow.
3. Return findings with file:line evidence.

checklist.md

Download checklist ↓
# Bilingual interface checklist
- English and Arabic communicate the same promise.
- Arabic uses RTL; code remains LTR.
- Buttons and headings do not overlap.
- Findings include evidence and verification.
03

Know when the starter is complete.

Trigger

It activates for the intended bilingual-review request and stays quiet for unrelated work.

Least privilege

It can inspect with Read, Grep and Glob but cannot edit or execute commands.

Evidence

Every finding names a real file and line, uncertainty, and a reproducible check.

Portfolio

Show the package tree, trigger cases and one anonymised review report.