How to translate an entire book with AI (Claude Code instead of ChatGPT) for a polished, consistent translation

For those who have never used Claude Code, even if the “black window where you type commands” (the terminal) intimidates you. No coding skills needed: just know how to type text.

1. If you’ve already translated a book with chatGPT or claude.ai, you know the struggle

Claude translates well, everyone agrees. So you open claude.ai, paste a passage, it gives you a nice translation, you copy it into Word. For a short excerpt, perfect. But translating an entire book that way, after a few chapters, it goes downhill.

First, the endless copy-paste. Each passage is a paste-and-copy routine. For a book of several hundred pages, that’s thousands of back-and-forths, not to mention botched pastes, forgotten passages, mixed-up order.

The clever ones think that instead of pasting bits, just throw the whole file .txt at once. But that creates other problems. One: the chat window has a length limit, a file too big isn’t fully taken into account, silently truncated without you knowing. Two: even if it swallows it, Claude doesn’t translate the whole book in one go; it translates a bit then suggests “continue for the rest,” and each “continue” is a chance to skip an entire paragraph or drift in tone from what came before. In the end, you get a long translation full of gaps, and cross-checking it against the original takes longer than translating yourself.

Worse, each conversation has its own memory. Today you tell it “this character uses informal speech, this term translates like this.” Tomorrow, new conversation, it’s forgotten everything. You have to repeat it all, every time fearing you’ll forget a rule set the day before.

Hence the worst mistake when translating a long book: loss of consistency. In chapter 2 a term is translated one way, in chapter 9 another, unintentionally. The tone, formal at the start, gradually dilutes into colloquial speech. A careful reader will feel the book derailing, while you can no longer remember what you decided over the weeks.

Finally, the longer the discussion, the more Claude runs out of steam. Stretch a conversation over ten chapters and toward the end, it gets confused, loses the thread, quality drops without you understanding why.

The problem isn’t that Claude translates poorly. It translates very well. The problem is that claude.ai isn’t built for a long-term project. Claude Code is.

2. What changes with Claude Code?

Imagine claude.ai as an excellent translator locked in a room with no window, with no book in sight. You have to read each passage aloud to it, then hand-copy what it dictates. And every time you call, it’s forgotten the previous translation.

Claude Code is the same translator, but sitting at your desk. The book is in front of them. Next to it, the glossary you’ve built, and a “translation rules” sheet pinned on the wall they re-read before picking up the pen each time.

The difference: Claude Code works directly on your computer’s files, not through a chat window. It opens chapter files itself, writes the translation into a file itself, consults the glossary you keep in the folder itself. Concretely:

	claude.ai (web)	Claude Code
Reading and writing files	No, you have to copy-paste	Reads the original, writes the translation on its own
See the whole book	No, only what you paste	Sees the whole folder: original, translation, glossary
Remembering rules from one time to the next	Each chat is an island, everything is forgotten	Re-reads the rules file each time
Self-checking	You have to proofread manually	Run some checks: missing sentences, broken formatting
Long conversation	The longer it goes on, the more muddled it gets	One chapter, one clean session, never overloaded

The three items in the right-hand column are the key to a translation that’s both polished and consistent: a brief file that Claude rereads before each translation, a glossary that keeps terms consistent, and an automatic proofreading pass that catches mistakes you’d otherwise miss even after an hour of checking.

3. Don’t be afraid of the terminal (the “black window”): setup in ten minutes

Claude Code runs in a terminal, that window where you type commands instead of clicking. It looks intimidating, but in reality, most of the time you’re typing in French like you’d send a message to tell it what to do. A real command sounds pretty ordinary, like:

Translate livre-complet.txt into French following the rules in regles-traduction.md, then save the result as livre-fr.txt.

Installation only needs to be done once:

1. Open the terminal. Windows: Start button, type Terminal, open it. Mac: open the Terminal.
2. Install Node.js (the foundation for running Claude Code): go to nodejs.org, download the LTS version, install it like any other software.
3. Install Claude Code : type this exact line and press Enter. npm install -g @anthropic-ai/claude-code
4. Start it up : navigate to the folder created specifically for the book translation, type claude --dangerously-skip-permissions The first time, it’ll guide you through logging in with your Claude account. That’s it.

If a scary red error line appears, no need to panic. Copy it and paste it back to Claude asking “how do I fix this error?”: it’s the best tool for troubleshooting itself.

Which subscription, for what budget

To use Claude Code, you need a paid Claude account; the cheapest is the Claude Pro subscription, around €17 per month. From experience, the €17 plan is enough to translate a book; no need to upgrade. The key thing to know: after a while, you’ll exhaust your session quota (the tokens); you just need towait about 4 hours for it to recharge, then type “continue”.

To give you an idea, for a book of about 300 pages: count on a total of fifteen to twenty hours spread out, but two-thirds of that is waiting time while the quota refills, not time spent staring at the screen; the actual time Claude is translating is around 5 to 7 hours. In other words, it’s the kind of task you kick off and then come back to every so often with a “continue,” not something you watch in one go. (Figures are indicative, depending on the book’s difficulty)

4. How to translate a book so it stays consistent

The secret isn’t in the “translate this for me,” but it’s also not about dictating every step in detail. The simplest approach, and the one best suited for beginners, is to give Claude a really complete prompt from the start, then let it plan and work. In practice, from start to finish, you’ll only need about two commands : a startup prompt (long, pasted once) and then the word “continue.” The heavy lifting, including building the framework, is Claude’s job; your role is mainly to validate and answer its questions.

Your part: a folder and a file

Create a folder for the book and drop the book file into it. That’s about all the manual work you have to do. In the end, the folder will look like this, but all files other than the book are created by Claude (you just review and correct as needed):

mon-livre/
  livre-complet.txt      ← the entire book, the ONLY thing you drop in
  source/                ← the original split into chapters (created by Claude)
  traduction/            ← the translation (created by Claude)
  CLAUDE.md              ← the logbook Claude re-reads each time (created by Claude)
  glossaire.md           ← the glossary (created by Claude)
  regles-traduction.md   ← style rules (created by Claude)
  exemple-traduction.md  ← a sample excerpt to set the tone (created by Claude)
  personnages.md         ← characters and relationships (for a novel, created by Claude)
  tu-vous.md             ← who uses formal/informal "you" (for a novel, created by Claude)

You don’t have to split the book into chapters either. Just ask Claude to translate: it reads as it goes and cuts the book into small portions itself, because a single chapter is often already too much text to swallow at once. To make checking and resuming easier, it’s better to have one file per chapter in source/, and again, Claude handles that.

The startup prompt (to paste once)

This is the most important command; it covers everything. Paste this block as-is into Claude Code, only modifying the part in brackets according to your book:

I want to translate the book from the file livre-complet.txt from [English] to French, with high quality and a consistent tone from start to finish. [This is the novel “Pride and Prejudice” by Jane Austen.]

PHASE 1, preparation (once finished, stop and ask me your questions, don’t translate yet):

1. Skim the book to grasp its genre, context, and tone: read the first chapters in full, then sample passages further in and the overall structure (no need to load everything into memory at once).
2. Ask me all the clarifying questions before starting (depending on the book: formal vs. informal “you” between characters, whether to keep footnotes, whether to translate or keep proper names, any other uncertain points, etc.).
3. Translate 1 or 2 representative passages in 2 or 3 different tones so I can choose; save the chosen version in exemple-traduction.md.
4. Propose a translation plan for the whole book (how to divide it, how far to go per session).
5. Build the framework yourself and save it in files (I’ll review and correct if needed):
• CLAUDE.md : the style, important translation rules, the workflow (including how to save as you go and how to resume after an interruption).
• glossaire.md : table of proper names, recurring terms and expressions, with a proposed translation.
• regles-traduction.md : the writing rules.
• Since it’s a novel, add personnages.md (the characters and their relationships) and tu-vous.md (who uses tu or vous with whom depending on the situation: everyday life, intimacy, tension, in front of others; decide based on context and era, I’ll correct it).

PHASE 2, Triggered only after my validation of phase 1. Every time I say ‘continue’, apply these rules exactly.

6. Translate the book portion by portion following the plan (usually one chapter at a time, or a finer breakdown if the chapter is too long). Save as you go in traduction/ after each passage, noting where you are; don’t wait until the end of the portion to save.
7. After each portion, compare yourself to the original and correct anything that’s a mistranslation, stiff, or lost a nuance.
8. During translation, as soon as a term, proper name, or new decision appears, immediately update the relevant .md files; if you need to change an already-used translation, also correct previous chapters to stay consistent, then let me know.
9. Every two or three portionss, review the entire folder traduction/ to spot terms, proper names, or tu/vous choices translated differently from one chapter to another; list them and harmonize according to glossaire.md , and tu-vous.md.
10. If the session is interrupted along the way (quota exhausted, computer off, window closed), next time I’ll simply say ‘continue’: you open the current file, see where you stopped, and resume exactly at that spot, without retranslating from the beginning or forgetting anything.

Now start with PHASE 1. Do steps 1 to 5, then STOP, ask your questions, wait for my validation.

It seems long, but you only paste it once. Two points make all the difference:

• Phase 1 forces Claude to stop, present a plan, and ask its questions before translating. It’s the best habit to adopt with Claude Code: it brings to light decisions you hadn’t thought of (do those two characters use tu or vous, do we keep the footnotes, are proper names anglicized or not) from the start, instead of letting a bad choice silently spread over three hundred pages.
• Phase 2 teaches Claude in advance to work over time : saving as it goes, proofreading, keeping the reference .md files up to date, checking overall consistency from time to time, and knowing how to resume alone after an interruption. So, afterwards, you just type ‘continue’.

Then, just say ‘continue’

Once the plan is validated and questions resolved, each time you want to move forward, simply type continue. Claude rereads the plan and files, sees where it is, and resumes exactly where it stopped. Even if the previous session was cut off mid-way (quota exhausted, machine off), ‘continue’ brings it back to the current line, without retranslating or forgetting anything, because the translation and progress are already in the files.

What the .md contains (Claude builds it, you can correct it)

The points below are just to help you understand what Claude creates; it’s not manual work for you.

CLAUDE.md is the logbook Claude rereads at each opening, where major decisions are recorded. A draft proposed by Claude will look like this; you just have to correct two or three lines:

# Translation Project: 'Pride and Prejudice' (Jane Austen)

## Style
French, literary and elegant register, keeping Austen's light irony.

## Intangible Rules
- Civilities (Mr/Mrs/Miss) are rendered consistently.
- Dialogues sound natural but remain polite, according to characters' rank.
- Austen's long sentences are aired out to flow in French, without chopping the meaning.

## Chapter-by-chapter progress
1. Read source/chapter-NN.txt
2. Translate according to regles-traduction.md, glossaire.md, tu-vous.md
3. Save in traduction/chapitre-NN.fr.txt
4. Proofread: no missed sentences, no leftover original text, tu/vous usage compliant

glossaire.md guarantees that a name or term is always translated the same way, from start to finish. Claude reads the whole book to list them; you just approve the translations you like:

Source	Retained translation	Note
Mr. Darcy	M. Darcy	keep ‘M.’, don’t anglicize the name
entailment	hereditary substitution	legal term, stay consistent
accomplished	accomplished (of a young lady)	DO NOT translate as ‘finished’

Next to it, regles-traduction.md record the writing rules (explicitly note puns in parentheses, never use dashes, etc.). These two files don’t need to be perfect from the start: you build them up as you go, just like in step 8 of the prompt.

exemple-traduction.md contains one or two translated passages as a model to lock in the ‘tone’, and it’s the only place where your taste is indispensable. Claude translates the same passage in several tones, you pick the one you like (or tweak it), and it becomes the reference:

## Original
It is a truth universally acknowledged, that a single man in
possession of a good fortune, must be in want of a wife.

## Desired translation
C'est une vérité universellement reconnue qu'un célibataire pourvu
d'une belle fortune doit avoir besoin d'une épouse.

## Why this way
- Keep Austen's understated, elegant irony.
- Preserve the ample rhythm of the original sentence.
- A polite register, faithful to 19th-century sociability.

For a novel: tu/vous forms and character relationships

French has its own coherence trap: the tu and vous. Two characters might address each other formally at first, then switch to informal once they become close; the choice depends on the relationship, rank, and era. Get the register wrong, and the whole atmosphere rings false. For a novel, Claude therefore builds personnages.md (who is who, what relationships) and tu-vous.md (who addresses whom with tu or vous, depending on the situation). Best: let Claude propose based on context, then adjust.

Speaker	Addressee	Situation	Form
Elizabeth	Darcy	beginning of novel, tense relationship	you (formal)
Elizabeth	Jane	sisters, close	you (informal)
Mrs. Bennet	her daughters	everyday	you (informal or formal, depending on era)

With these two tables, the tu/vous distinctions hold together across the whole book, without wavering from chapter to chapter.

Consistency across the entire book, not just chapter by chapter

This is where web apps fall short and Claude Code excels, because it can re-read all the already translated chapters. Step 9 of the startup prompt is exactly that: every 2-3 chapters, Claude scans the folder traduction/, spots where chapter 3 translated something one way and chapter 17 another, and harmonizes. A task impossible to do by hand over hundreds of pages, and here it’s just one line in the prompt. That’s what ‘homogeneous from start to finish’ really means.

One chapter, one session, and don’t be afraid of /clear

Claude has a limited short-term memory, called the ‘context’. If you translate a dozen chapters in the same conversation, by the eighth it gets saturated, confused, and quality drops. That’s the ‘the longer it goes, the more it runs out of steam’ from earlier.

A common question: if I type /clear, does Claude lose everything? Yes and no. /clear clears the current conversation’s memory, but that’s exactly what you want to clean so it stays sharp. It doesn’t lose anythingbecause the plan, the .md files, and the translation are on the hard drive; the new session just needs to re-read them to rebuild all the context. After each chapter, do /clear, then type ‘continue’ for the next one. The rule is one sentence: one chapter, one session. Thus the last chapter is as good as the first.

5. When you run short mid-translation: picking up again without losing anything

Translating a big book spans several days, sometimes weeks; inevitably, at some point, your quota runs out (Claude shows a message like “You’ve hit your session limit…” then stops), or your computer shuts down, the network drops, you close the window by accident. With Claude Code, you lose almost nothing.

The reason comes down to one point: everything essential is already on the hard drive, not in the session’s temporary memory. The translation is saved in traduction/, the glossary in its file, the major decisions in CLAUDE.md. The conversation is just a temporary workshop; losing it doesn’t destroy the work. The trick of “saving as you go” is already built into the startup prompt (steps 6 and 10): Claude writes each passage as soon as it’s translated, so if it’s cut off at passage 40 out of 100, the first 39 are safe.

The procedure is ultra-simple: if it’s the quota, wait a bit then reopen (the limit recharges in cycles); the work stays intact on the disk, nothing to do. When you return, just type “continue”, and Claude reopens the current file and picks up right where it left off, without retranslating or drifting.

On a long translation project, I even let it run on its own overnight: the machine chained parts one after another, and as soon as it hit the token limit, it paused for an hour and then tried again, always restarting from the last saved line. Over the whole duration, no sentence was translated twice or forgotten, thanks to a single principle: save as you go, resume from the file.

6. The recipe in brief

1. Create a folder, drop the entire book in it (preferably a .txt file). That’s all your manual work.
2. Paste the startup prompt (long, complete). Claude reads the whole book, builds the .md framework, tries a few tones as tests, then stops to present its plan and ask you its questions.
3. Validate the plan, answer the questions, set the tone via the sample excerpt. That’s about the only real decision you have to make.
4. Then, just type “continue”. Claude translates chapter by chapter, saves as it goes, self-reviews against the original, and keeps the reference .md files updated continuously.
5. Overall consistency (sweeping through all traduction/ every x chapters) and resumption after interruption are already in the prompt: Claude handles it alone.
6. Do /clear between chapters to keep it sharp; no fear of losing anything, everything is in the files.

For a thick book, using Claude Code (instead of the web interface) is what separates “translated to get it done” from “translated for good.” You don’t need to be technical, nor to write a whole set of rules. You just need to know how you want it to sound, let Claude propose the rest and say yes or no, and let the tool maintain the discipline for you, from the first to the last page.