What this covers
Pasting 3000 words into ChatGPT and saying “translate this to Japanese” produces something fluent, grammatical, and tonally wrong half the time — branded terms get translated, the author’s voice gets flattened, and idioms become Wikipedia-sounding paraphrases. This guide is the workflow I use when the output has to be publishable, not just intelligible: build a glossary first, paste source plus glossary, instruct on tone, then verify with a back-translation diff.
Who this is for
Anyone whose translation will be read by a native speaker who cares — a marketer localizing a newsletter, a researcher publishing in a second language, a founder writing investor updates for a non-English board. If the audience just needs the gist, paste-and-pray is fine. If they’ll spot a clumsy verb choice or a brand name rendered as a generic noun, you need a workflow.
When to reach for it
- Localizing marketing copy where tone and brand terms matter.
- Translating a long-form piece (1000-5000 words) and preserving a specific authorial voice.
- Producing bilingual documentation where terminology has to stay consistent across files.
- Handling domain text (legal, medical, technical) where wrong terms have real cost.
Before you start
- Pick a model. GPT-5 for legal or technical material where reasoning matters; GPT-5.5 for general copy; GPT-5.4 only for low-stakes drafts.
- Decide which terms must NOT be translated (brand names, product SKUs, code identifiers, person names). Write them down.
- Identify the tone you want preserved — punchy, formal, conversational, marketing-glossy — and find a 100-word sample of that tone in the target language to anchor on.
- Open a Project for the language pair so memory and glossary stay scoped — don’t pollute your general chat.
Step by step
-
Build a glossary from the source. Paste 500-1000 words and ask:
Extract every term in this text that needs a deliberate translation choice: brand names, product names, technical terms, idioms, recurring phrases. Output as a table: source term | suggested translation | reason | leave-as-is (Y/N). -
Review the glossary by hand. Lock brand names as leave-as-is. Pick a single rendering for each repeated technical term. This is the step that prevents inconsistency on page 6.
-
Translate one section at a time (500-1500 words per turn), pasting the locked glossary at the top:
Translate the section below from EN to JA. Glossary (use exactly): [paste table] Tone: conversational, like a founder writing to early users. Preserve paragraph breaks. Do not summarize or shorten. -
Run a back-translation check on suspicious paragraphs:
Translate this JA paragraph back to EN literally, preserving structure. Don't smooth it out — I want to see drift.Compare to the original. Drift on factual nouns or numbers is a real problem. Drift on adjectives is usually fine.
-
For idioms or culturally loaded phrases, ask for 3 alternatives with a one-line explanation each, then pick:
Give me 3 ways to render "we shipped it warts and all" in JA marketing copy, each with a one-line note on the connotation it carries. -
Final pass: a native speaker reads it. The workflow gets you 90% of the way; the last 10% is human judgment.
A prompt that produces honest output
You are translating EN to {target language} for a published piece.
Constraints:
- Use the glossary I pasted. Don't invent new renderings for terms in it.
- Preserve the author's voice: [paste 100-word voice sample].
- If a phrase is ambiguous in source, flag it inline as [AMBIGUOUS: ...]
rather than guessing.
- Do not paraphrase to make sentences shorter. Match length where possible.
- Numbers, dates, and proper nouns must round-trip unchanged.This catches the cases where the model would have silently smoothed something into a half-meaning.
Quality check
- Spot-check 3 random paragraphs by back-translating literally. Look at factual nouns and numbers, not stylistic shifts.
- Run a find-replace pass on glossary terms — every locked term should appear in the locked rendering, no exceptions.
- Read the first and last paragraphs out loud in the target language. If they sound like a translation, the voice prompt didn’t take.
- Cross-check brand names, URLs, and product SKUs survived intact — these are the most common silent edits.
How to reuse this workflow
- Save the per-language Project with its glossary; reuse it for every piece into that language pair.
- Keep a
translation-voice-samples/folder with the 100-word anchor texts for each tone you commonly target. - For team use, share the glossary as a CSV so engineering, marketing, and support all use the same rendering.
Recommended workflow
Glossary extraction → human review → section-by-section translation with glossary pinned → back-translation diff on suspicious paragraphs → idiom alternatives → native-speaker final pass.
Common mistakes
- Pasting the whole document at once. Long passes lose tone consistency by the end; the model drifts toward generic.
- Skipping the glossary step. Brand names get translated, technical terms get rendered three different ways across the doc.
- Asking for “natural” without anchoring with a voice sample. “Natural” defaults to the model’s house style in that language, which is bland.
- Trusting back-translation to be lossless. It isn’t — but drift on factual content is a red flag worth investigating.
- Using a fast model for legal or medical text. The stakes don’t match the tradeoff.
- Forgetting that ChatGPT’s training in low-resource languages (e.g. some Southeast Asian or African languages) is weaker — quality drops, and back-translation drift increases. Get a human reviewer earlier.
FAQ
- Is ChatGPT better than DeepL for translation?: DeepL is often cleaner for short, conventional text. ChatGPT wins when you need glossary control, tone preservation, or domain instructions. Use DeepL for first-pass, ChatGPT for the polish.
- Can I translate code comments and docs the same way?: Yes, but pin a glossary that explicitly says “leave code identifiers, variable names, and string literals unchanged.” Code-adjacent prose is where translators go wrong most often.
- Does the model preserve markdown / HTML?: Mostly, but it occasionally drops a tag or rebalances headings. Diff against source after translation.
- What about voice mode for translation?: Useful for travel-grade conversation. Not useful for publishable output — there’s no glossary, no back-translation, no audit.