Imported from isArman/scientific-fa-translation-skill (SKILL.md). Install upstream with npx skills add isArman/scientific-fa-translation-skill. Copyright stays with the author.

Scientific Persian translation

English → academic Persian for papers, articles, books, and technical documentation. Accuracy, consistent terminology, and print-ready RTL outrank literary fluency. Cursor chat is not the RTL surface.

When to use

• Translate a scientific or technical document into Persian.
• Review a finished Persian translation against these rules (references/review.md).
• User mentions ترجمه علمی, راست‌چین, RTL, PDF, چاپ, paper, article, book.

Not for: writing or explaining code, commit messages, PRs, UI copy, literary or marketing translation, casual Persian conversation.

Locked defaults

Override only when the user says so.

Workflow

1. Preflight. scripts/preflight.sh — know which engine and fonts exist before promising a build. Confirm source, target, and level. Jobs and subject are inferred from the source in step 3, not chosen from a list.
2. Ingest. references/source-ingest.md: fetch the source, extract figures, run scripts/crop-source-figures.py, flatten with scripts/prepare-figures.py figures/, then scripts/prepare-figures.py figures/ --check. Write inventory.md and manifest.txt in the working tree. Never translate from memory when a fetch fails. Never ship a black pdfimages dump. Never ship a full source-page raster as a figure. A book contents / brief contents page is inventory chrome, not optional: translate it and print it.
3. Terminology first. Read enough of the source to name three jobs (practices the source actually covers: DevOps, networking, Linux, …) and the subject (the product or protocol: nginx, …). Announce all four with the level. Keep those jobs' and that subject's lexicon English (location, proxy_pass, deployment — not مکان / گذرگاه پیش‌رو / استقرار). Ordinary prose stays Persian. Apply references/terminology.md. Do not write glossary.local.md and do not append to glossary.md. Write terms.tsv in the working tree (discarded with the job) before drafting — keep-English rows must include forbidden_fa. Show close calls to the user first. Lint always with --terms terms.tsv.
4. Read references/scientific-style.md and references/rtl-bidi.md. For anything past ~15 pages also references/long-documents.md.
5. Translate with the ensemble in references/ensemble.md: Composer and Grok draft, Luna judges diffs only. Do not add, omit, or soften claims; preserve hedges (may, might, suggest, remain unknown).
6. Isolate every LTR run in the print source — whole clusters, one isolate each (references/rtl-bidi.md).
7. Lint. scripts/check-fa.py doc.tex --level <level> --terms terms.tsv --manifest manifest.txt --strict and clear every error. --strict will not run without those two files. Lint each part as you finish it, not at the end. --pairs FILE adds rows; it does not replace term-pairs.tsv.
8. Build and verify. scripts/build-pdf.sh doc.tex <slug> --verify lints again, checks figures/ when that directory exists, and will not copy a PDF if lint, figure check, or --verify fail. --verify also rejects an HTML-engine PDF when XeLaTeX is installed (copy-paste would reverse Persian). Look at the rasterised pages. Run the judgement checklist below.

If the user asks for HTML only, use assets/rtl-document.html. If they ask for Markdown, wrap the body in <div lang="fa" dir="rtl">, still isolate LTR spans, and say print RTL will be weaker than PDF. Reverse translation (FA→EN) only on explicit request; then drop the RTL rules and keep technical terms in English.

Terminology in one paragraph

Full policy and the field-term test: references/terminology.md. Infer three jobs and one subject from the source; there is no domain pack. Ordered, first match wins: generic document chrome (Abstract, Figure) is always Persian; named artifacts and acronyms are English; a 2–5 word technical label is English as one whole isolate; the inferred subject's lexicon is English at every level (nginx, location, proxy_pass); the jobs' lexicon is English at system-docs (deployment, reverse proxy) and Persian at journal unless it is also the subject. Everything else is Persian. Never half-translate (خوشه Kubernetes, بلوک location). A kept-term plural is the singular stem plus ها (\en{service}ها, \en{location}ها, \en{API}ها), never services / locations / APIs; ezafe on Latin is still forbidden (Goی). Never mix two forms of one concept in a document. Forbidden calques are enforced from references/term-pairs.tsv plus this job's terms.tsv.

Example (system-docs, jobs DevOps / networking / Linux, subject nginx): «برای \en{proxy_pass} در هر \en{location} از یک \en{upstream} استفاده کنید.» — not «برای گذرگاه پیش‌رو در هر مکان».

Example (journal): «این پیاده‌سازی از \en{gradient descent} برای کمینه کردن تابع هزینه استفاده می‌کند.»

Persian mechanics

Full rules: references/scientific-style.md. UTF-8; ک not ك, ی not ي; نیم‌فاصله in می‌شود, می‌توان, نمی‌کند, داده‌ها; punctuation ، ؛ ؟ «». Formal verb forms only (register is judgement). Letters, ZWNJ on the listed verbs and plurals, Western digits, Latin punct, and unisolated number clusters are machine-checked.

RTL

Full rules: references/rtl-bidi.md; engines and measured limits: references/pdf-output.md. Chat does not need to be RTL.

On the PDF, non-negotiable: isolate every English term, whole collocation, number cluster, formula, URL, and inline code with \lr{…} (or <span dir="ltr">). Slash-, space-, arrow-, or parenthesis-joined English (OP_IF/OP_NOTIF, 3.1 The OpenStack service, STARTED -> LOCKED_IN, 1.0.1 (2026-08-09)) is one isolate — split across two spans it renders reversed on the page. Listings are LTR and left-aligned. Math stays LTR. Do not mirror images. After a trailing English insertion the Persian period must belong to the Persian sentence.

Output

Default: printable PDF at /home/$USER/Documents/books/<slug>.pdf ($HOME/Documents/books, created if needed). Report that absolute path, the page count, and the engine used.

1. Read references/pdf-output.md. Prefer assets/rtl-document.tex.
2. Persian prose in the .tex; English runs in \lr{…} / \en{…}.
3. Listings in \begin{latin}…\end{latin}. Captions translated, identifiers kept (Figure 3 → شکل 3).
4. \includegraphics each cropped artwork file inside LTR; HTML <img dir="ltr"> with margin-inline: auto. Flatten alpha. Order, aspect, and subfigure layout preserved. Never point at a full pdftoppm of the source page.
5. Bibliography in a latin section, source language. Fill the colophon with source, licence, and retrieval date.
6. scripts/build-pdf.sh path/to/doc.tex <slug> --verify. Without TeX the same script takes the filled-in assets/rtl-document.html; embed Vazirmatn with scripts/fetch-vazirmatn.sh and never the UI-FD cut. That HTML PDF displays correctly but copy-paste reverses Persian — say so. A page-range PDF is scripts/extract-pdf-pages.py in.pdf out.pdf 1-20 — one range, never a per-page loop (references/pdf-output.md).

Quality gate

Machine-checked — scripts/check-fa.py --level <level> --terms terms.tsv --manifest manifest.txt --strict must exit 0. It covers orthography (ک/ی, نیم‌فاصله on listed verbs/plurals, Western digits, Persian punctuation), forbidden calques at that level, half-translated noun phrases, English -s plurals of kept terms, leftover Latin ezafe (Goی), split isolates, un-isolated Latin runs, un-isolated number clusters (ranges and dates reverse on an RTL page), listing direction, mirrored artwork, missing images, figure direction, full-page figure rasters, and terminology drift inside isolates. Do not re-check these by hand. --pairs and --terms merge onto the house list.

Judgement — only these five, and they are the whole point:

• No added, omitted, or softened scientific claim; hedges intact
• Terminology consistent: one form per concept; inferred jobs and subject lexicon stayed English; consistent with terms.tsv
• Every source figure present, unmirrored, in source order, with a translated caption, showing the artwork (not a black box, not a dump of the English source page around it)
• Rasterised pages actually read correctly (periods, parentheses, listings, no missing-glyph boxes) — not judged from pdftotext
• Claim-changing ambiguities were asked, not guessed; the rest are reported

Delivery — final PDF at /home/$USER/Documents/books/<slug>.pdf, chat is a short pointer with the path, page count, engine, and queued questions.