Most writing tools try to fix text after it’s written; this project treats AI-like, flat, and templated Japanese as a generation-time problem plus deterministic detection. The core insight: prevent predictable LLM hand‑prints by locking in a head-skeleton and style constraints up front, then mechanically surface remaining issues for human/agent judgment instead of blind wholesale replacement.
What Sets It Apart
-
Generation-time constraints plus mechanical linting: a 12-article "writing constitution" constrains outputs before they are produced, and a sudachipy-based lint (lint.py) deterministically finds templated or translation‑y patterns. So what: fewer post-hoc rewrites and clearer prompts for agents.
-
Pragmatic diagnostic modes: quick/full/exp scoring returns a 0–100 "naturalness" score (machine base + human-adjustment bands) and actionable top reasons without changing the document. So what: teams can audit AI drafts before committing to edits or publishing.
-
Autonomous rewriting loop and six-axis rubrics: the v1.5.0 self-edit harness scores drafts on axes like脱AI臭 (de-AI-ification), information density, and warmth, then iterates until thresholds are met. So what: preserves voice and context while reducing formulaic phrasing, rather than performing blind replacements.
Who it's for and trade-offs
Great fit if you regularly use LLMs to draft business Japanese (minutes-to-thousands-of-words reports, meeting minutes, guides, blog posts) and need reproducible checks: editors, in-house comms, engineering docs writers, and agent builders integrating writing skills. Look elsewhere if your priority is raw translation/localization, non-Japanese languages, or if you need an out-of-the-box GUI: this is a skills/CLI-first toolkit that expects integration into agent workflows and some configuration.
Additional notes
The repository bundles scripts (lint.py, outline.py, terms.py), an opt-in reading-load detector for heavy cognitive load points, and guidance for genre-specific thresholds. It targets agent ecosystems (Claude Code and similar) and is MIT-licensed; the project emphasizes opinionated, corpus‑validated rules rather than heuristic word lists.