thkt/kousei

/kousei - Japanese Proofreading

Proofread and auto-fix Japanese text using textlint.

Input

• $1: file path / #123 (Issue) / #pr123 (PR) / raw text
• If empty → target the most recently edited .md file

Language Support

| Language | textlint (mechanical) | Structure review (AI) | | -------- | --------------------- | --------------------- | | Japanese | Yes | Yes | | English | Skip | Yes |

Detection: if text contains ≥50 hiragana/katakana/kanji characters → Japanese. Otherwise → English (structure review only).

Execution

| Step | Action | | ---- | --------------------------------------------------------------- | | 1 | Determine target | | 2 | Get original content | | 3 | Run bun x textlint --fix (Japanese only) | | 4 | Structure review (AI-powered, see checklist below, all languages) | | 5 | Show textlint diff + structure feedback | | 6 | For Issue/PR → present changes, confirm before push |

Target Resolution

| Input Pattern | Action | | ------------------ | ----------------------------------------------------- | | *.md file path | Run textlint --fix on the file | | #N (Issue number) | gh issue view N --json body → fix → gh issue edit | | #prN (PR number) | gh pr view N --json body → fix → gh pr edit | | Other text | Write to temp file → fix → display result | | Empty | git diff --name-only to find recent .md files |

textlint Execution

cd ~/.claude/textlint && bun x textlint --fix --config .textlintrc.json "$file"

Structure Review (Anti-Work-Slop)

After textlint fixes, review the text against these structural quality checks:

| Check | Question | | ------------------ | ---------------------------------------------------------------- | | Author's judgment | Is the author's own conclusion stated in 1-3 lines before AI output? AI output should be supporting material, not the main content. | | Half-able | Can this be cut in half? If not, the author may not know what matters. | | Fact vs opinion | Are facts, assumptions, and proposals clearly labeled and separated? | | Action clarity | Is the expected reader action explicit? ("Review X" not "ご確認ください") | | Reader cost | Does this reduce the reader's cognitive load, or shift work to them? |

Report only items that need improvement. If structure is sound, skip silently.

Format:

Structure review:
- [Check name]: [specific suggestion]

Anti-AI-Pattern Review (Japanese)

After structure review, check for AI-typical Japanese patterns:

| Pattern | Signal | Fix | | --- | --- | --- | | Appended significance | 「〜を浮き彫りにしており」「〜に示唆を与えています」 | Data speaks for itself. Cut the appendage | | Roundabout copula | 「〜として位置づけられています」「〜の役割を果たしています」 | 「〜です」 | | Hedge stacking | 「〜と考えられます」「〜の可能性があります」 in succession | Commit or hedge once | | Boilerplate conclusion | 「今後の展開が注目されます」「〜が期待されます」 | Say something specific or end without it | | Connector stacking | 「さらに」「加えて」「また」 in consecutive sentences | Vary or cut. One per 3 paragraphs max | | Full-width dash | 「——」(U+2014 x2) | Replace with parentheses or comma |

Apply to Japanese text only. Report format:

Anti-AI-pattern review:
- [Pattern]: [specific location and fix]

Issue/PR Update Flow

1. Fetch body → write to temp .md file
2. Run textlint --fix
3. Show diff
4. Ask user: "Apply these changes?"
5. On approval → gh issue edit N --body "..." or gh pr edit N --body "..."

Display Format

When changes found:

Proofread complete: N fixes applied

[diff of changes]

When no issues:

Proofread complete: no issues found