Claude Code in Persian: the language setting, CLAUDE.md, and right-to-left text

Yes, for the answers. The official settings reference states that there is no fixed list for response languages: the value goes to the model verbatim and Claude Code does not validate it, so persian and farsi both work. Two other places work differently: the app interface language is its own fixed list, and voice dictation has a fixed list too; Persian is not on the dictation list, so choosing it falls back to English dictation. The text responses are unaffected.

Put { "language": "persian" } in ~/.claude/settings.json; farsi works as well. The official example is japanese, which shows the type is any language name; the value is not validated and a typo raises no error, so check the spelling yourself. The project-level .claude/settings.json can carry the same key, and to make a whole repository consistent, also state the convention in ./CLAUDE.md.

Keep two effects apart. The Arabic script is written from right to left, and conformant implementations must reorder it for display with the Unicode Bidirectional Algorithm, so the screen runs right to left while the file is still stored left to right; in a mixed line, digits and English runs still read left to right. When the algorithm needs an extra hint, W3C recommends the isolate controls RLI, LRI, and PDI rather than RLE, LRE, and PDF. Editors have official switches: Vim's rightleft makes a window display right-to-left, and GNU Emacs has bidi-display-reordering.

The half-space is U+200C ZERO WIDTH NON-JOINER. Its width is zero, so a character that seems to have vanished usually has no glyph but is still there. The real risk is deletion: chapter 9 of the Unicode Standard states that ignoring or removing these non-joiners produces text with a different meaning or meaningless text, and the Persian plural suffix is the official example. Turn on your editor's display of invisible characters when you debug; W3C also notes that control codes are invisible and easy to leave unterminated, and that plain text can only use them.

Yes. CLAUDE.md is a plain-text file that Claude reads at the start of every session, so write the rules in the language your team reads. Length and placement matter more than language: stay under 200 lines, because longer files reduce adherence; block-level HTML comments are stripped before injection, so do not hide rules in them; and when the file grows, use @path imports for the rest.

That is a team convention, not a vendor fact, but write it down either way. A common split is Persian for explanations and English for code and commit messages: Persian letters inside identifiers bring two concrete problems. Bidirectional scripts raise official security considerations when used in identifiers, and Persian ی and ک are different code points from Arabic ي and ك, so search and code review miss them easily. Put the rule in CLAUDE.md and it is read at the start of every session.