Glossary and translation quality

Two levers to make translations better for your specific book.

Glossary

Every translate-intent project has a Glossary — a list of source → target term overrides Claude respects while translating. Use it for:

• Proper nouns (character names, place names, product names) that

the model would otherwise transliterate inconsistently

• Technical vocabulary specific to your book (a physics manual's

variable notation, a manga's fictional taxonomy)

• Preferred phrasings ("kimono" over "traditional Japanese garment")

Add entries from the Glossary tab in the project header. You can also paste bulk entries as source → target per line.

Entries apply to future translations only. Rerun affected pages to apply glossary changes to already-translated content.

Quality scores

Every completed page gets a 0-100 score from the quality reviewer. Under 70 (configurable per project) triggers a Claude refinement pass that rewrites the translation with more care and flags it in the UI as "Refined by Claude" with a summary of what changed.

Reviewers look for:

• Untranslated Japanese runs slipping through
• Sentence-boundary drops
• Chart / table labels that got left behind
• Diagram callouts pointing at the wrong element
• Voice / tone drift within a paragraph

You'll see these called out on the Review tab of each page with click-to-jump excerpts and suggested fixes.

Making a NEEDS_REVIEW page shippable

• Read the Review tab — usually the reviewer already tells you what's

wrong.

• Fix the source scan if it's a quality issue (re-scan → replace

image on the page).

• If a term is consistently wrong, add it to the Glossary and

reprocess.

• Otherwise reprocess the page (may pick a different route on the

next run).