How to Use Claude for Translation: Long Documents, Glossaries, and Review

Claude is a strong tool for translating chapters, articles, and sections of a long document, and its newest models can read an entire book in one conversation. The mistake is to assume that because Claude can read the whole book, it will translate the whole book consistently. Three documented behaviors get in the way: a cap on how much Claude writes in one response, Projects switching to search-based retrieval once your files get large, and long chats being summarized to make room.

None of these is a flaw. They are how Claude stays usable with large inputs. But each one decides where your glossary should live and how big each translation request should be. This guide covers the Claude apps (claude.ai, desktop, and mobile) and the Claude API. Every limit below comes from Anthropic's own documentation, checked on 2026-10-12. Anthropic updates these pages often, so check the links before you plan around a number.

What Claude Accepts and Returns

Sources: Upload files to Claude, context window on paid plans, API PDF support, and the models overview.

Two details in those pages matter for translation more than the headline numbers:

• Only PDFs are read visually, and only short ones. In the apps, Claude analyzes both text and images, charts, and graphics in PDFs of 100 pages or fewer. From 101 to 1,000 pages it processes text only. For DOCX, EPUB, and other non-PDF files, Anthropic says Claude "extracts text only" and cannot read embedded images. If you upload a 300-page illustrated PDF or an EPUB, assume Claude sees the words, not the layout, captions in images, or charts.
• Claude can return a file, but it is a new file. With code execution and file creation turned on, Claude can create Word documents, PDFs, spreadsheets, and slides that you download. That is useful for a short translated document. It is still a document Claude builds by writing code, not your original EPUB or PDF with the language changed. Expect to check headings, footnotes, images, and page structure yourself.

Three Claude Behaviors That Break Long Translations

1. The output cap, not the context window, sets your chunk size

Anthropic lists a 1M-token context window and a 128K-token maximum output for every current API model. Anthropic's own rule of thumb is that 1M tokens is roughly 555,000 English words on the current tokenizer. By the same ratio, 128K output tokens is roughly 71,000 English words. That is our arithmetic from Anthropic's published figures, not a measured test, and the real number is lower if your target language uses more tokens per word or if you ask for notes alongside the translation.

So a novel fits in what Claude reads but not in what it writes back in one response. Translation answers are as long as their source, unlike summaries. Pick the chunk size from the output you need, not from the input Claude can hold. One chapter, or a few thousand words, keeps each response far below the cap and keeps each review unit small enough that a missing paragraph is easy to spot.

2. A large Project stops reading your whole book

Projects let you upload files once and reuse them across chats. On paid plans, Anthropic says that "when your project knowledge approaches the context window limit, Claude will automatically enable RAG mode." In RAG mode, Claude does not load every document. It searches the project and retrieves the parts most relevant to your request (RAG for projects). You cannot turn this on or off yourself.

For questions about a book, that is a sensible design. For translation consistency, it changes what you can rely on. If you upload the whole book, earlier translated chapters, and a glossary file, then ask for chapter 14 "consistent with the earlier chapters," Claude may retrieve only some of that material. This is our inference from Anthropic's description, not a documented failure rate. It means you should not trust a large Project to remember how you translated a character's nickname in chapter 2.

3. Long chats get summarized

On paid plans with code execution turned on, Anthropic says Claude "summarizes earlier messages to make room for new content" when a conversation nears the context limit (automatic context management). Your chat history is kept, and the chat can keep going. But a glossary you pasted in message one may be working from a summary by message forty. Anthropic also notes that these longer conversations use more of your usage limit.

Where Your Glossary Should Live

This table follows from the three behaviors above. It is the main decision a Claude translation workflow needs to get right.

The rule is simple: anything the translation must obey goes into the request itself or the instructions, not into a file Claude might not retrieve. Use project knowledge for material Claude may look things up in, not for rules.

A Working Method for a Long Document

This is a method, not a benchmark result. It turns the behaviors Anthropic documents into steps you can repeat and audit. Adjust the chunk size to your plan and model. The terminology consistency guide covers what belongs in a glossary.

1. Freeze the source and number the chunks. Work from one final file. Split on chapters or sections, not pages, so sentences and footnotes stay whole. Label them C01, C02, and so on, and note each chunk's first and last sentence.
2. Build the glossary in its own chat. Upload the full book and ask Claude to list recurring names, places, terms, and forms of address, with a proposed translation and the page where each first appears. This is where the large context window earns its keep: Claude reads everything once. Have a qualified person approve the list.
3. Set up the project. Put the short style brief and the must-never-change terms in project instructions. Upload the source file as knowledge if you want Claude to look up context, but keep the full glossary out of knowledge.
4. Use one chat per chunk or a few chunks. Start each request with the full approved glossary and the chunk. Ask Claude to begin with the chunk ID and end with END C07, or the matching ID, so a response that stopped early is obvious. A fresh chat avoids summarization creeping into a thread that has run for hours.
5. Check every chunk before moving on. Compare paragraph counts and confirm the last source sentence has a translation. If the END marker is missing, request the chunk again rather than continuing from the middle. Add any new term to the glossary, approve it, and include it in later requests.
6. Search the assembled translation for drift. Look up every glossary term and every source-language name. Variant spellings and leftover source words are the usual signs.
7. Review risky passages with a person. Names, numbers, dates, negation, and anything legal, medical, or headed for publication need a qualified reader.

If you use the API

The same method scripts well, and two API features help with cost:

• Prompt caching lets you reuse a fixed prefix, such as the system prompt with the full glossary and style brief, across requests. Anthropic prices cache reads at a fraction of the base input price (5% on Opus 5.5 and Sonnet 5.5, per the models overview).
• The Message Batches API costs 50% less than standard requests, and Opus 5.5, Sonnet 5.5, and Haiku 5.5 support up to 300K output tokens per batch request with a beta header. Batches suit chunk translation because chunks rarely depend on each other once the glossary is fixed.

Neither feature changes the output: the API returns text. You still need to rebuild the translated document from it.

A Prompt That Survives Chunking

Use the same four-block structure as our ChatGPT translation guide and Gemini translation guide, with chunk control added. Claude's own prompting guidance favors clear, explicit instructions, and tagged sections make the boundaries obvious:

<task>
Translate chunk C07 from [source language] to [target language].
</task>

<audience>
[Who will read it and what they need to do with it.]
</audience>

<rules>
- Apply the glossary exactly. If a term is missing, keep the source term in
  [brackets] and list it after the translation as NEW TERM.
- Preserve names, numbers, dates, citations, and paragraph boundaries.
- Do not summarize, skip, or add explanations.
- Start with "C07". End with "END C07".
</rules>

<glossary>
[full approved term list]
</glossary>

<source>
[chunk text]
</source>

Keep review separate from translation. Once a chunk is done, open a new request with the source chunk, the translation, the glossary, and a short rubric: omissions, additions, changed numbers or names, glossary violations, and wrong register. Treat Claude's findings as things to check, not verdicts. Like any model, it can miss errors or flag correct choices. For fiction, voice drift between chapters is a separate risk from terminology drift. The AI novel translation workflow covers character bibles and voice tests.

Check Privacy Before You Upload

What happens to your document depends on your account type:

• Free, Pro, and Max: whether your chats are used to train Anthropic's models is controlled by the "Help Improve our AI models" setting under Settings > Privacy (model improvement settings). If you allow it, Anthropic may keep data from new or resumed chats in de-identified form for up to five years (data retention). If you don't, Anthropic says its existing 30-day retention period continues (consumer terms update). Incognito chats are not used to improve Claude, even with the setting on (model training). Chats flagged by safety classifiers can still be used for Anthropic's trust and safety work.
• Team, Enterprise, and the API: Anthropic says the consumer training settings "do not apply to services under our Commercial Terms," which include Claude for Work and API use. Check your organization's agreement for its retention terms.

For contracts, manuscripts under NDA, or client files, settle this before you upload anything. The document translation privacy checklist lists the questions to ask any tool, including ours.

When Claude Is the Right Tool, and When It Isn't

Which model is strongest for your language pair is a separate question. Our best LLM for translation guide explains how to test that rather than trusting a leaderboard.

Claude is the right tool when you can see and fix every failure: a chapter you will read closely, a glossary you are drafting, or a translation you are reviewing. It is the wrong main tool when the job is a complete file. At that point you are managing chunks, glossaries, end markers, and layout by hand. That is the work BookTranslator is built for. You upload the PDF, EPUB, DOCX, or other supported file, it translates the whole document with context and a consistent glossary, and you get a translated file back, plus a bilingual version where the format supports it. For scanned PDFs, OCR mode rebuilds a clean translated PDF from the recognized content rather than copying the original page layout exactly. You see the price before checkout, so you can upload the file and decide from there. Whichever tool you use, keep the review steps above for anything important.