An agent skill that scrubs machine-translation odor from Japanese prose

What it does

natural-japanese is an Agent Skill for Claude (and compatible agents) that refines Japanese work documents—minutes, reports, blogs, essays—into clear prose. It structures writing into three stages: argument design before drafting, a 12-article style constitution that constrains generation, and post-hoc mechanical inspection via lint.py using sudachipy morphological analysis. The tool flags forbidden phrases, monotonous rhythm, and English-calque syntax such as inanimate subjects with transitive verbs, then loops until the text converges.

The interesting bit

The project assumes LLMs cannot reliably smell their own robotic tics, so detection is strictly deterministic—sudachipy parses text and matches patterns—while the actual rewriting is left to the agent or human. For subtler readability issues like comma placement or subject-verb distance, where the authors found mechanical thresholds fail in corpus testing, the skill falls back to AI-driven review against antipattern catalogs rather than faking certainty with regex.

Key highlights

• Prevents AI odor at the source by constraining generation with a 12-article “style constitution” instead of only rewriting after the fact.
• lint.py uses sudachipy to deterministically flag forbidden phrases, translation-calque syntax, and rhythmic monotony.
• Iterates in a closed loop: fix, re-lint, diff against --baseline, and repeat until findings converge.
• Optional --reading-load lane points out structural readability issues—long sentences, buried lists, kanji chains, double negatives—without pretending to score them.
• Ships with document-type templates in doctypes/ for minutes, research reports, slide decks, and essays.

Caveats

• The experimental semantic.py topic-flatness detector depends on torch and sentence-transformers and downloads a roughly 1 GB model, so it is strictly opt-in.
• Mechanical linting deliberately does not judge word order, comma placement, or subject-verb distance; the authors’ corpus tests showed automated thresholds are unreliable in that space, so those still require AI or human review.

Verdict

Grab it if you generate Japanese business prose with Claude and are tired of hand-editing machine-translation odor. Skip it if you write only English or prefer to keep editorial judgment entirely manual.