Prompt Traduzir + Tom

Translate with tone: prompts that preserve voice across languages

A literal translation often loses what made the original work in the first place — the playful Twitter casualness of “vc curtiu?” flattens into a flat “did you like it?”, an empathetic support reply turns corporate, a punchy marketing CTA becomes a polite suggestion. This generator builds a ChatGPT or Claude prompt that translates a passage while explicitly steering the tone: formal business, casual social, academic, marketing or empathetic customer support.

The prompt is structured around the RTCF framework so the LLM has every signal it needs to make the right linguistic choice instead of guessing.

Role: Translator with tone-shifting expertise
Task: Translate {text} from {source} to {target}
Context: Audience is {audience}, formality {level},
         brand voice {voice}
Format: Output translation + brief tone notes

Five tone presets that cover most use cases

• Formal business — LinkedIn outreach, investor updates, contracts.
• Casual social — Twitter/X, Instagram captions, TikTok scripts.
• Academic — journal articles, dissertations, conference abstracts.
• Marketing — landing-page CTAs, e-mail subject lines, ad copy.
• Empathetic customer support — refund flows, apology e-mails.

Why LLMs beat raw machine translation

Google Translate and even DeepL are excellent at semantic equivalence but limited at tone. Modern LLMs such as GPT-4 Turbo, Claude 3.5 Sonnet and Claude 3 Opus can hold an entire brand voice in context and adjust register on demand. The trade-off: cost and latency. A 1000-word passage costs roughly $0.001 on the API — orders of magnitude cheaper than a human translator at $0.20+/word, but slower and pricier than DeepL’s flat-rate API.

Cross-cultural traps to flag in your prompt

Brazilian Portuguese is famously direct; the same line translated word-by-word can sound abrupt in British English or curt in Japanese. Spanish has the tú/usted split, French uses tu/vous, Brazilian PT swings between você and the regional tu. Always pin the audience (country, age, formality, industry) inside the Context block, otherwise the model picks a default that may be off-key. For high-stakes copy — legal, medical, top-of-funnel marketing — pair the LLM draft with a human reviewer.

Localization platforms vs. raw prompts

For multi-string products, dedicated platforms still win: Lokalise, Crowdin, Phrase and localize.com handle locale files, translation memory, glossary terms and reviewer workflows. The raw-prompt approach shines for one-off pieces — a single blog post, a marketing e-mail, a customer-support reply — where the overhead of plugging into a TMS is bigger than the benefit.

FAQ

How do I encode cultural sensitivity? Spell it out inside Context: target country, age bracket, formality expected, taboos to avoid. The model cannot infer what your audience finds offensive.

What about highly technical content? LLMs translate medical, legal and engineering text well, but a domain reviewer is still mandatory for liability reasons — the model can confidently invent a wrong term.

How much does the API cost? Roughly $0.001 per 1 000 words on GPT-4 Turbo or Claude 3.5 Sonnet, depending on input vs. output mix. Long passages benefit from caching.

Can I version the prompt? Yes — store it in a .prompt file under Git, or use LangSmith, PromptLayer or Helicone to A/B test variations against a regression suite.