A linter for AI-generated Korean prose

What it does

im-not-ai is a Claude Code (and Codex CLI) skill that rewrites AI-generated Korean text to sound human. It leaves every fact, number, and quote untouched, but scrubs translation-ese, mechanical parallelism, hedging overload, and emoji abuse. Think of it as a linter for prose voice rather than code syntax.

The interesting bit

The project treats “AI smell” as a taxonomy, not a vibe. It catalogs 60+ sub-patterns across 10 categories (translation artifacts, hedging, rhythm uniformity, visual decoration, etc.), assigns severity tiers (S1/S2/S3), and runs them through a multi-agent pipeline: detect → surgically rewrite → audit for fidelity → review for naturalness. The strict mode even grades output A–D and will force a human review if it over-edits.

Key highlights

• Two modes: Fast (single call, ≤5,000 chars, ~3 min) vs. strict (5-agent pipeline, 8,000+ chars auto-promoted)
• Four hard rules: meaning preservation, span-only edits, genre maintenance, and a 50% change-rate kill switch
• Install paths: Claude Code marketplace plugin, git clone + ./install.sh, or Codex CLI (fast mode only)
• Artifacts: saves final.md plus intermediate JSONs (detection, fidelity audit, naturalness review) in dated workspace folders
• Community ports: unofficial web UI at im-not-ai-ocx.illuwa.click, plus a separate Codex desktop adapter

Caveats

• Strict 5-agent pipeline is Claude Code only; Codex gets fast mode
• Requires Claude Code CLI — won’t run in web Claude or standard ChatGPT
• README notes the project is a “skill” (prompt/agent harness), not a standalone model or traditional library

Verdict

Worth a look if you publish Korean content edited by LLMs and want to stop sounding like a translated whitepaper. Skip it if you don’t use Claude Code or Codex CLI, or if your output language isn’t Korean — the pattern taxonomy is specifically built for Korean translation artifacts and honorific drift.