1.9K ·GitHub
What is this?
An Agent Skill for writing and revising clear, natural Japanese for work documents — meeting minutes, reports, internal guides, proposals, blog posts. Its whole point is stripping the telltale 'AI smell' from machine-written prose.
An Agent Skill for writing and revising clear, natural Japanese for work documents — meeting minutes, reports, internal guides, proposals, blog posts. Its whole point is stripping the telltale 'AI smell' from machine-written prose.Is this a standalone app, or something that lives inside my AI coding agent?Not an app you open: it installs into an AI agent's skill directory (Claude Code, Cursor, ChatGPT, Codex) and is invoked automatically during everyday writing requests. The lint scripts can also run standalone in a terminal.
Not an app you open: it installs into an AI agent's skill directory (Claude Code, Cursor, ChatGPT, Codex) and is invoked automatically during everyday writing requests. The lint scripts can also run standalone in a terminal.Which AI agents can use it — do I have to be on Claude Code?Claude Code is the primary target, but the openskills route makes it usable from Cursor, ChatGPT, Codex and similar agents through AGENTS.md. Running the lint scripts standalone only needs uv, which resolves each script's dependencies automatically.
Claude Code is the primary target, but the openskills route makes it usable from Cursor, ChatGPT, Codex and similar agents through AGENTS.md. Running the lint scripts standalone only needs uv, which resolves each script's dependencies automatically.Does the automatic detection catch every unnatural pattern?No — corpus validation (180+ comparative analyses) showed that word order and comma placement are hard to judge mechanically, so those spots rely on rule-catalog manual review plus an auxiliary reading-load check (--reading-load). Flagged issues are triaged, never bulk-replaced.
No — corpus validation (180+ comparative analyses) showed that word order and comma placement are hard to judge mechanically, so those spots rely on rule-catalog manual review plus an auxiliary reading-load check (--reading-load). Flagged issues are triaged, never bulk-replaced.How does it actually keep AI text from sounding AI-written?Prevention over repair: before writing, the agent fixes a heading skeleton and applies a 12-rule style constitution ('write the conclusion first', 'don't repeat the same pattern'). Leftover habits are flagged mechanically by lint.py using sudachipy morphological analysis, while a human or the agent decides what to actually change.
Prevention over repair: before writing, the agent fixes a heading skeleton and applies a 12-rule style constitution ('write the conclusion first', 'don't repeat the same pattern'). Leftover habits are flagged mechanically by lint.py using sudachipy morphological analysis, while a human or the agent decides what to actually change.What can I ask my agent to do once the skill is installed?Drafting or proofreading minutes, reports, internal guides and proposals; instructions like 'make it readable', 'write from the conclusion', 'fix unnatural phrasing'; revising AI-written drafts; diagnosing AI-smell with a score; and drafting or rewriting blog posts and essays.
Drafting or proofreading minutes, reports, internal guides and proposals; instructions like 'make it readable', 'write from the conclusion', 'fix unnatural phrasing'; revising AI-written drafts; diagnosing AI-smell with a score; and drafting or rewriting blog posts and essays.What happens after the linter flags issues?The agent does not mass-replace flagged phrases — that would flatten nuance. It sorts the findings and runs an autonomous revision loop against a 6-axis revision rubric (de-AI-smell, information density, warmth and more) until the text reaches the passing bar.
The agent does not mass-replace flagged phrases — that would flatten nuance. It sorts the findings and runs an autonomous revision loop against a 6-axis revision rubric (de-AI-smell, information density, warmth and more) until the text reaches the passing bar.Can it measure how artificial a document sounds without rewriting it?Yes — a diagnostic command scores a document's naturalness from 0 to 100 without touching the text, handy for checking how strong the AI smell in an existing draft is.
Yes — a diagnostic command scores a document's naturalness from 0 to 100 without touching the text, handy for checking how strong the AI smell in an existing draft is.What are similar projects?yomiyasu, im-not-ai, linkedin-agent-skill, asd-ste100-skill, humanizer, Humanizer-zh, clarity, no-ai-slop
yomiyasu, im-not-ai, linkedin-agent-skill, asd-ste100-skill, humanizer, Humanizer-zh, clarity, no-ai-slopWhich topics is it in?lint, ai-writing, japanese, writing-tools, text-processing, writing, claude-skills, nlp, documentation, claude, agent-skills, claude-code
lint, ai-writing, japanese, writing-tools, text-processing, writing, claude-skills, nlp, documentation, claude, agent-skills, claude-code