Skip to main content

Read

Translate a whole book in one run, without the terminology drifting

Any AI assistant will translate a paragraph well. None of them will translate a book — you would be pasting chapters in by hand for a week, and getting back four different renderings of the same term. Start this once and it works through all 300 pages holding one agreed glossary, so chapter nine still uses the words chapter one settled on.

Rare

One run for the whole book, not one paragraph at a time

Sentence quality stopped being the hard part a while ago; length is. Feed a 300-page work through a translator in pieces and each piece is done with no memory of the last — the same term comes out four ways, and by chapter six you are reading forty translations stapled together. Here the whole book is one run, held to one glossary.

  • You start it once. It reads the document, agrees the terminology with you, then translates every chapter against that same glossary — no re-pasting, no per-chapter babysitting.
  • Progress is checkpointed. Close the app mid-book, come back, and it resumes from where it stopped rather than starting the bill over.
  • Sections are checked for the failures that are mechanical — dropped content, numbers that changed, terms that broke the glossary — and sent back for another pass before you ever see them.
Pasted in one piece at a time
Ch. 1注意力头Ch. 3关注头Ch. 6注意头Ch. 9注意力单元
“attention head”, four ways — each piece had no memory of the last
1Glossary, agreed onceYou confirm · the only stop
  • attention head注意力头
  • embedding嵌入
  • TransformerTransformerkept
Every chapter held to the same list
2The whole book, start once
  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
  7. 7
  8. 8
  9. 9
  10. 10
  11. 11
  12. 12
  • Failed a check (a number changed), so it went back for a second pass
  • App closed here; it resumes at chapter 8, not chapter 1
3One term, one rendering, cover to cover
Ch. 1注意力头Ch. 9注意力头

One glossary, agreed before a word is translated

Drift is a bookkeeping problem more than a language one, so it is solved with bookkeeping, not a better model: fix each term once, up front. That is where the one decision you have to make sits.

  • It reads the document first and works out what it is — subject, register, audience — then builds the list of terms that will recur.
  • You see that list and correct it. This is the only checkpoint that needs you; everything after it runs.
  • If your folder already contains translations that settled on a rendering, that choice carries forward instead of being re-litigated.

What happens after you press start

  1. It reads the document and writes a brief

    Subject, register, audience, and the terms that will recur. Nothing is translated yet.

  2. You correct the glossary

    The one place your judgement is worth more than another pass of compute.

  3. Every chapter is translated against it

    Same terms throughout, including terms inherited from translations already in your folder.

  4. It checks itself and reworks failures

    Dropped content, changed numbers and broken glossary terms go back for another pass automatically.

  5. You read it against the original

    Translation beside source text, aligned, so checking a passage does not mean hunting for it.

  6. You polish chapter by chapter

    Polish arrives as a tracked change you accept or reject, the same as everywhere else in the editor.

The rest of it

PDF, EPUB, Word, Markdown

The formats long documents actually arrive in.

Output is yours

Markdown in your own folder, next to the original. No export step, no second copy to file.

Terms inherited across documents

A rendering settled in one book carries into the next one in the same folder.

Polish as tracked changes

Chapter-level rewrites arrive as marked-up edits you accept or reject, not as silent replacements.

Interruption is survivable

Checkpointed progress, so a closed laptop is not a wasted run.

Mechanical checks before you read

Missing content and changed numbers are caught by the machine, so your attention goes to judgement instead.

Where it gets hard

  • A real book costs real money

    A 300-page work is a large job. Translate the first chapter and look at it before committing the rest.

  • Poetry, wordplay and legal text

    Terminology consistency is exactly the wrong tool for writing that depends on ambiguity. Fine for a textbook; treat literary translation as a draft to argue with.

  • Scanned sources need recognition first

    Recognition quality sets the ceiling for everything downstream — do that step first and check it.

  • You still have to read it

    The verification pass catches mechanical failures, not bad judgement. It cannot tell you a sentence is technically correct and completely wrong in tone.

Common questions

Can it translate an entire book, not just a page?

Yes. You start it once and it works through the whole document chapter by chapter. Progress is checkpointed, so closing the app mid-book resumes rather than restarting.

How does it keep terminology consistent across hundreds of pages?

It reads the document first and builds a glossary of recurring terms, which you can correct before translation starts. Every chapter is then translated against that same glossary.

Where does the translation end up?

As Markdown in your own folder, next to the original — not as an export you have to download and file somewhere.

Try it on one chapter before you commit a book.

Credits are spent on the translating. Reading and editing the result is not what you are paying for.