DEV Community

Artem KK
Artem KK

Posted on • Originally published at kazkozdev.hashnode.dev

Why Book Translation Needs a Second Pass

Why Book Translation Needs a Second Pass

Most LLM translation demos stop after a single generation pass. That is enough to preserve rough meaning, but not enough to preserve rhythm, tone, and narrative continuity across long chapters.

Book Translator uses a two-step workflow:

  1. Draft translation for semantic fidelity.
  2. Self-reflection pass for style, flow, and readability.

That extra pass matters because long-form translation quality breaks down in subtle ways. Literal phrasing accumulates. Transitional sentences become stiff. Paragraph rhythm starts sounding machine-generated even when each sentence is technically correct.

The project treats translation less like one-shot prompting and more like an editorial pipeline. It runs locally with Ollama, which keeps sensitive manuscripts off third-party APIs while still giving you a repeatable CLI workflow.

Key design choices:

  • chunking for long documents
  • local-first inference via Ollama
  • explicit self-reflection stage for refinement
  • CLI-first workflow for repeatable runs

If you are building long-form AI writing systems, the main lesson is simple: generation quality is often a workflow problem, not just a model problem.

Repo: https://github.com/KazKozDev/book-translator

Top comments (0)