RTL writing (Hebrew, Arabic, Persian, Urdu with English terms)

Purpose

Mixed RTL + English text renders by one rule: each line (paragraph, bullet, heading, table cell, bold label) takes its direction from its first letter. Start a line with an English word, number, symbol, emoji, link or inline code and the whole line flips left-to-right and its punctuation jumps. So every line opens with an RTL word. Never translate technical terms: names, commands and terms of art stay in English inside the sentence. Once invoked, this applies to every later reply until the user says stop. Examples are Hebrew; for Arabic, Persian or Urdu also read references/arabic-persian-urdu.md once.

Chat replies — write once, no tool call

Write the reply directly: no reference to read, no checker to run, no draft to retype. The rules below are writing habits, applied as each line is written: do not plan the reply line by line in your thinking and do not review it line by line afterwards; think about the content as you would without this skill.

1. Hebrew opener on every line. "Python is…" → "השפה Python היא…"; a name → "החברה Marvel…", "הסרט Iron Man…"; "malloc returns" → "הפונקציה malloc מחזירה"; "3 steps" → "שלושה שלבים". Prefixes work too: "ב-Docker", "ה-API", "ל-Python".
2. Headings, bullets, bold labels, table cells follow rule 1, header cells included: "## הגדרת Docker", "- הפונקציה fork יוצרת…", "הכלי Podman: …", "| הכלי Docker | נפוץ |". Labels that feel like fixed English terms are the usual slip: "Pros:" → "יתרונות:", "Step 1:" → "שלב 1:", "Train:" → "סט ה-Train:".
3. After a bold label that ends in English, the next word is Hebrew: "סשן ה-Transformer: מודל…".
4. End lines on a Hebrew word, not on English + punctuation: "…ומריצים את make בטרמינל.". If no Hebrew ending fits, drop the final period.
5. Numbers next to English: "Iron Man (2008)" → "הסרט Iron Man משנת 2008"; "-1" → "מינוס 1".
6. Code blocks hold no Hebrew (comments and strings in English); the line right after a block opens in Hebrew.
7. Last resort for a cell or line that is only a name: prefix the letter א with no space ("אDocker").

The one second look worth taking: headings and bold labels, where most slips happen ("## Replication ו-FSMO" → "## שכפול ו-FSMO roles"). If the user reports a line that rendered wrong, fix it; the full rule set (below) covers rarer cases.

Files you write

Files are checked mechanically, on the file itself (nothing to retype):

• Markdown: python3 scripts/check_rtl.py FILE --fix → fix until 0 errors (stdin works: -).
• Word: set w:bidi / w:rtl (references/docx.md); python3 scripts/check_docx_rtl.py FILE → 0 paragraphs without RTL.
• LaTeX: references/latex.md (polyglossia, XeLaTeX). PowerPoint: references/pptx.md (rtl="1").
• Course terminology for a document: references/glossary/<family>.md (first mention: Hebrew term with the English in parentheses); a COURSE.md may override terms.

Files in this skill

• scripts/check_rtl.py — checks Markdown/chat text for every RTL script; --fix suggests openers; --json.
• scripts/check_docx_rtl.py — Hebrew/Arabic paragraphs in a .docx that lack w:bidi.
• references/chat-markdown.md — the full rule set with every edge case (tables, links, lists, mixed quotes).
• references/docx.md, references/latex.md, references/pptx.md, references/arabic-persian-urdu.md, references/glossary/.

When this triggers

The user invokes it (/rtl-writing, the old /hebrew-with-english, or a namespaced form); complains that RTL lines come out flipped or punctuation jumps; asks for an RTL document or slides with English terms; or another skill is about to write RTL prose into a file. Not for translation, grammar questions, or plain RTL prose with no English tokens.