epoko77-ai/im-not-ai

What it solves

This project addresses the "AI-like" quality of Korean text generated by LLMs (such as ChatGPT, Claude, and Gemini). It specifically targets translationese, mechanical structures (e.g., "First, Second, Third"), overused AI idioms, excessive passive voice, and repetitive sentence rhythms that make AI-generated Korean feel unnatural to native speakers.

How it works

The tool operates as a skill for Claude Code and OpenAI Codex CLI, using a sophisticated taxonomy of 70+ sub-patterns across 10 categories to detect and rewrite AI-typical expressions. It employs a dynamic routing system called route_hint to determine the processing path based on the text's quality:

• Light (1 LLM call): For well-written text with minimal AI markers; performs a conservative rewrite.
• Standard (2 LLM calls): For typical AI drafts; includes a diagnosis phase followed by a targeted rewrite.
• Heavy (3+ LLM calls): For severe "AI slop," very long documents (over 15,000 characters), or when strict verification is needed; includes diagnosis, rewriting (with parallel chunking if necessary), and a finalization phase.

To ensure fidelity, the system follows four strict rules: it preserves meaning (facts, figures, proper nouns), modifies only detected spans, maintains the original genre, and prevents over-editing (warning at 30% change rate, stopping at 50%).

Who it’s for

• Writers and editors who want to remove the "AI smell" from Korean drafts without changing the core meaning.
• Users of Claude Code or Codex CLI who need a professional, natural-sounding Korean output.
• People dealing with translation-style Korean that often results from English-centric AI models.

Highlights

• Korean-Specific Taxonomy: Based on academic translation research, targeting specific patterns like excessive use of "through" (~를 통해) or specific double particles.
• Efficient Routing: Reduces token costs and time by using a single call for high-quality text instead of a heavy pipeline.
• Multi-Agent Architecture: Uses specialized agents for diagnosis (humanize-diagnostician), monolithic rewriting (humanize-monolith), and final verification (humanize-finalizer).
• Quantitative Metrics: Integrates z-score based metrics (e.g., comma usage patterns) to objectively measure and verify the "human-like" quality of the text.