TroubleshootingAnswered

Does Claude's translation get more paraphrased toward the end—is that normal?

Anonymous7/16/2026· Views 109

I'm having Claude translate a novel chapter by chapter. The beginning comes out pretty literal, but toward the end it starts summarizing whole sentences or adding expressions that aren't in the original. I'm not sure if it's because the context gets too long or if it's just random, but can a prompt fix this, or is it just something I have to accept as part of how the model works?

Answers 1

You need to log in to write an answer.

첫째StaffLv.67/16/2026

Hey, it's the admins here. Bottom line: this is actually a normal characteristic of how Claude works, and it's a problem you can control quite a bit through your prompts. First, it helps to understand what's happening. As conversations (context) get longer, Claude tends to lose focus on earlier instructions and details from the original text. Anthropic actually calls this "context rot" in their official documentation about context windows. We don't have exact public numbers on how many chapters or tokens before it kicks in—it varies by model and situation—but the fact that longer conversations drift toward paraphrasing rather than literal translation isn't random; it's more of a structural characteristic. I'd recommend three practical ways to tackle this: 1. Start a new conversation (or new project conversation) for each chapter instead of piling everything into one thread. Don't keep the entire previous chapter stacked in context; instead, summarize just the terms and character details you need and feed those back. It's much more stable. 2. At the start of each chapter's prompt, explicitly spell out rules again like "sentence-by-sentence literal translation, no summarization or paraphrasing, no adding anything not in the original." The longer a conversation gets, the weaker the earlier instructions become, so restating them each time actually helps. 3. Don't scatter your glossaries, character settings, and style guides throughout the conversation—lock them in as knowledge in the Projects feature instead. Then whenever you start a new conversation, they'll be referenced consistently each time. So while there's some give-and-take with how the model works, just by keeping conversations shorter and making it a habit to restate the rules each time, you can cut down the paraphrasing problem you see pretty significantly. We've put together tips specifically tailored for longer projects like webnovel translations where you need to keep settings consistent across multiple chapters: https://usingclaude.com/ko/guides/features/claude-webnovel-character This guide also covers translation methods and their limits if you want to dig deeper: https://usingclaude.com/ko/guides/features/claude-translation-guide