translation-style-matchlisted
Install: claude install-skill rodar-dzn/translation-style-match
# Translation style matching
Make a new translation read as though the same hand produced it, by
measuring an existing translation instead of guessing at it.
## Scope
Use on text the user has the right to produce: their own work, licensed or
commissioned work, public domain, or in-house localization. If the goal is a
complete unauthorized edition of a copyrighted book, say so plainly once and
offer the audit and style-guide work instead.
Reference corpora stay on the user's machine under `corpus/`, which is
gitignored. Never commit corpus material, and never paste long stretches of
it into chat — quote the short spans a finding rests on.
## The four phases
Run them in order. Each produces a file the later phases depend on.
### Phase 1 — Profile
Turn the reference translation into a written, checkable description.
1. Convert the user's reference files to plain text:
```bash
python scripts/extract_corpus.py corpus/reference.epub --out work/ref/
```
Handles EPUB, FB2, and plain text. One file per chapter, spine order
preserved.
2. Measure it:
```bash
python scripts/fingerprint.py work/ref/ --out work/ref-fingerprint.json
```
3. Read `references/axes.md`. Fill `templates/STYLEGUIDE.template.md` by
sampling the corpus for each of the six axes. Every claim in the style
guide must cite a span you actually read. A style guide written from
assumption is worse than none — it launders guesses into rules.
4. Read `references/targets/<lang>.md`