Seo translator

SEO Translator

Battle-tested translation methodology from the SkillProof localization pipeline — the same rules used to localize 4,500+ pages into 10 languages. Produces native-sounding text, keeps SEO value, and burns as few tokens as possible.

Step 1 — classify the job (silently, before translating)

• SEO content (page titles, meta descriptions, headings, blog posts, landing pages): keywords must survive translation in the form people actually SEARCH in the target market — not the literal dictionary form. "Claude skills installieren", not "Fähigkeiten installieren".
• Marketing copy: register and idiom matter more than word fidelity. Translate the intent.
• Technical/UI: consistency and precision. Build a mini-glossary first if strings repeat.

If the source has <title>, meta descriptions, H1/H2 markdown, or the user mentions SEO/search — treat as SEO content without asking.

Step 2 — register per market (defaults that professionals use)

Language	Register	Like
German	informal "du" for dev/consumer tools, "Sie" for finance/legal	Linear, Stripe DE
French	professional "vous"	Stripe FR
Spanish, Italian, Portuguese	informal "tú"/informal	Notion ES
Dutch, Swedish, Danish, Norwegian	informal "je"/"du" — standard in tech	—
Czech, Polish	polite "vy"/neutral informal for dev tools	local SaaS
Japanese	です/ます form unless told otherwise	—

Ask only if the domain makes the default risky (medical, legal, government).

Step 3 — translate with these hard rules

1. Never translate: product names, brand names, code blocks, commands, file paths, URLs, email addresses, placeholders like {n} or %s, established tech terms the market keeps in English (commit, deploy, frontmatter, prompt).
2. SEO fields respect limits: titles ≤60 chars, meta descriptions ≤155 chars — rewrite to fit, don't truncate mid-thought.
3. Keywords: identify the 1-3 target keywords in the source (or ask if ambiguous). Render each as the target market searches it — check: would a native type this into Google? Keep keyword in title and first paragraph, as in the source.
4. Numbers, scores, dates: unchanged. Localize only formats (1,000 → 1 000 where native).
5. Links: URLs untouched; anchor text translated.
6. Structure: markdown/HTML structure survives exactly — same headings, same tables, same emphasis. If the source has a CTA block, translate only its visible text.

Step 4 — the anti-AI pass (mandatory)

Translated text fails if it reads like machine output. After drafting, re-read and fix:

• calques from the source language (word order copied, not natural)
• AI-tell vocabulary in the target language (the local equivalents of "delve", "crucial", "landscape", "seamless")
• uniform sentence length — vary it the way natives write
• literal idioms — replace with the target market's own idiom or drop
• title case in languages that use sentence case (most European languages)

One question per text: "would a native copywriter sign this?" If not, redo the failing sentence.

Step 5 — token discipline

• Translate in ONE pass; don't restate or summarize the source first.
• For batches: process sequentially in one response, no per-item preamble.
• Output ONLY the translation (plus a one-line note if a decision needs flagging). No "Here is your translation".
• For repeated strings across a batch, translate once and reuse verbatim.

Output format

Single text → the translation, nothing else. Multiple languages → one ## <lang> section per language. Batch of files/strings → same keys/structure as input (JSON in → JSON out, valid).

Do not

• Add translator notes unless a real ambiguity forced a choice
• "Improve" the source content beyond translation (unless asked)
• Translate quoted user testimonials' names or company names
• Trigger for code refactoring or spoken-language interpretation requests