translate

❌ Route elsewhere: new content → content:article / content:blog / content:social / content:draft; editing source-language content → content:editor; auditing it → content:audit.

Required (ask once; don't proceed without them): target language with locale variant (es-MX not es, pt-BR not pt, zh-Hans/zh-Hant, en-US/en-GB, de-DE/de-AT/de-CH, fr-FR/fr-CA) and register (formal / informal / neutral — drives T-V choice, honorifics, contractions). Source language: auto-detect; confirm only if ambiguous (Norwegian Bokmål vs Nynorsk, Serbian Cyrillic vs Latin). Optional: purpose (marketing / documentation / legal / UI strings / prose — sets how aggressively to localize), glossary, output destination.

Delivery rule — output goes EXACTLY where the user said:

• Named a file path → write there (overwrite only if explicitly invited)
• "in-place" / "replace" / a source file with replacement intent → overwrite the source
• Multiple locales into one file or structured format → the shape they described
• No destination → inline in this conversation; never invent a filename
• File input, no destination → inline AND offer to write it to a path

Phase 1 — Intake

In parallel with detecting the source language, remember(["translation glossary", "brand voice", "locale conventions", "prior choices"]). Glossary referenced but not provided → ask for the file or the entries.

Phase 2 — Comprehend the source

Before translating a word, fix the domain and its terminology conventions, register (T-V, honorifics, contractions, slang allowance), voice (rhythm, sentence length, persona quirks — transfers but adapts), audience (cultural knowledge that won't transfer), intent (inform / persuade / entertain / warn / instruct), and structure (markdown, HTML, code, placeholders, URLs, identifiers — SACRED). Then list every translation-risky item per the content-translate risk inventory.

Phase 3 — Plan locale strategy (decide once, apply consistently)

• T-V form (tu/vous, du/Sie, tú/usted, ty/vy, 你/您) by register, held throughout
• Plurals via the target's CLDR categories in ICU syntax, never an if-1 branch
• Inclusive/gendered language per target-locale convention, chosen by audience
• Dates and numbers to the locale norm (keep ISO if the source enforces it)
• Currency keeps target formatting, converted only on explicit ask (rates change); units go imperial→metric only for a non-technical purpose with user permission
• Cultural references localized for marketing/literary, preserved+glossed for technical/legal
• Warn when length expansion will break layout Load content-translate for the locale reference (formats, punctuation, expansion figures, plural counts, inclusive-language norms).

Phase 4 — Translate in passes, never sentence by sentence in isolation

SENSE (the natural target equivalent, freed from source syntax; read it aloud mentally) → STRUCTURE (restore exact markdown/HTML/code/placeholder positions) → CONSISTENCY (terminology, T-V form, register, voice locked across the text) → LOCALE (dates, numbers, currencies, units, punctuation, quotation marks to target).

Hard preserves, character for character: code fences and contents; inline code; placeholders ({name}, %s, %d, {{var}}, ${var}, <%= %>, ICU {count, plural, …}); URLs, file paths, identifiers; numbers inside code or specs (only display numbers get locale-formatted); HTML tags and attribute keys; JSON/YAML keys (translate values only unless the schema permits).

Phase 5 — Self-review

Back-translate 2–3 critical sentences and fix drift; check one rendering per concept, one register and T-V form, and placeholder, code-block, and link counts against the source; rewrite anything that reads like source syntax or a dictionary lookup.

Phase 6 — Deliver

Per the Delivery rule. Then memorize() durable terminology choices, per-locale brand-voice rules, T-V decisions, recurring untranslatables (importance 0.7–0.9).

Files beyond a safe single pass (roughly 2000+ words): load content-translate and batch by its rules, carrying the glossary and T-V decision across batches.

VALIDATE:

🔍 Translation review (<source-lang> → <target-locale>) · <verdict>

Verdict: ✅ ship / 🟡 minor fixes / 🔴 needs rework
Overall accuracy: <N>/10 · Naturalness: <N>/10 · Terminology consistency: <N>/10 · Structure integrity: <pass/fail>

Findings (severity-ordered)
- 🔴|🟡|🟢 <one-line issue> — source: "<quote>" → translation: "<quote>" — <why it's wrong / fix direction>

What's working
- <one-line strength>

Then offer: Want me to apply the fixes? Reply with "apply all" / "apply critical" / "apply <#>".

NATIVE-AUDIT: the content-locale-humanize report shape, plus the same offer.

ADJUST: the revised translation in the original's shape, handed back to the user (not a file unless they named one), then content-translate's adjustment delta note.

Translation notes: after the translation, only when non-trivial choices were made, one bullet per choice in content-translate's notes shape.

N target locales → N independent blocks, each with its own Notes, no cross-references. Source already in the target language → say so, don't translate. The ask is generation, not translation ("translate this idea into Spanish marketing copy") → route to content:article or content:blog with a one-line explanation.