epoko77-ai/im-not-ai

AI가 쓴 한글을 사람 글처럼 윤문하는 Claude 스킬 — Korean AI-text humanizer: detects and rewrites translationese, mechanical parallelism, and 71 other AI tells

What it solves

This project addresses the "AI-like" quality of Korean text generated by LLMs (such as ChatGPT, Claude, and Gemini). It targets specific patterns common in AI-generated Korean, such as translationese (English-style phrasing), excessive use of English terms, mechanical structures (e.g., "First, Second, Third"), and repetitive AI-specific idioms, transforming them into natural, human-like Korean without altering the original meaning, facts, or figures.

How it works

The tool operates as a CLI skill compatible with Claude Code, GitHub Copilot CLI, and Codex CLI. It uses a tiered processing architecture based on a route_hint to optimize cost and speed:

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

It employs a taxonomy of 10 categories and 85 sub-patterns to detect AI markers and uses a "structure convergence gate" to verify that the output has actually shifted toward human-like distributions without over-editing (limiting change rates to 30-50%).

Who it’s for

  • Writers and professionals who use AI to draft Korean content but want it to sound natural and human.
  • Developers integrating AI-humanization workflows into CLI-based agent environments.
  • Users of Claude Code, GitHub Copilot CLI, or Codex CLI looking for specialized Korean linguistic refinement.

Highlights

  • Meaning Preservation: Strictly preserves facts, claims, numbers, proper nouns, and direct quotes.
  • Multi-Runtime Support: Officially supports Claude Code, GitHub Copilot CLI, Codex, and Gemini CLI.
  • Tiered Orchestration: Automatically selects the most efficient processing path (Light/Standard/Heavy) to reduce token costs.
  • Extensive Taxonomy: Uses a detailed system of 85 sub-patterns categorized by severity (S1-S3) to surgically remove AI markers.
  • Rigorous Verification: Includes gates to prevent over-editing and ensure the original author's voice is not erased.

Related

  • Project
    harshaneel/humanize

    humanize is an MIT‑licensed repo that ships two static, rule‑based skills—`humanize` (rewrites AI text to sound human) and `ai-check` (produces a forensic AI‑detection report). The skills work with any LLM agent (Claude Code, Codex CLI, ChatGPT desktop, Gemini, etc.) via a simple install script or marketplace command. Benchmarked on 25 diverse inputs, both internal and external perplexity‑based detectors rate the rewritten output as human‑like. The approach is offline, zero‑dependency, but cannot evade learned classifiers like GPTZero.

  • Project
    snflkd/fluent-korean

    A set of output-style guidelines and plugins for Claude and other LLMs to prevent broken machine-Korean and ensure clear, natural, and unambiguous Korean communication.

  • Project
    op7418/Humanizer-zh

    Humanizer‑zh is a Claude skill that rewrites Chinese prose to remove filler and repetitive phrasing while preserving facts, certainty, and the author’s voice. Install via `npx skills add …` or by cloning into `~/.claude/skills/`. Invoke with `/humanizer-zh` on inline text or files. It follows a 31‑point rule set (categories A‑F) and includes a test suite. MIT‑licensed.

  • Project
    Jakeschincariol/linkedin-agent-skill

    LinkedIn‑agent‑skill is a set of Claude prompt‑skills plus two tiny Python scripts that generate, clean, and score LinkedIn posts, comments, profile rewrites, weekly plans, carousel slides, DMs, inbox triage, and audits. It never posts automatically; you copy‑paste the final draft. The humanizer removes invisible characters, normalises punctuation, swaps buzzwords, and runs five local heuristics to give a “human‑score”. Install by copying the `li‑*` folders into `~/.claude/skills/` and optionally filling a `voice.md` file with your writing style.

  • Project
    devswha/patina

    A deterministic, pattern-based humanizer for multiple languages that removes AI-sounding phrasing from drafts while strictly preserving the original meaning.