im-not-ai: Open-Source CLI Agent Skill for Natural Korean AI Text Humanization

As generative AI models become essential fixtures in modern writing and workplace documentation workflows, drafting speed has accelerated dramatically, yet polishing away repetitive machine cadence and awkward translation tone remains a labor-intensive manual chore. The open-source agent skill 'im-not-ai' (plugin name: humanize-korean), hosted under epoko77-ai/im-not-ai, delivers a targeted development tool for terminal-based AI agent environments like Claude Code and Codex CLI that strips away robotic tells while faithfully safeguarding the author's original factual meaning and core structure.

Image source: @BBBang9900 (X / 꼬마봇)

Detecting Over 70 'AI Tells' with Fact-Preserving Rewriting

Korean text generated by leading foundation models such as ChatGPT, Claude, and Gemini often appears grammatically sound at first glance, yet frequently leaves readers with an unmistakable impression of mechanical composition.

At its core, 'im-not-ai' targets these subtle robotic traits by systematically classifying and detecting over 70 distinct varieties of 'AI tells' through refined prompt architecture and structural transformation patterns.

• Mitigating Translated Syntax: It eliminates awkward syntactic artifacts commonly found in direct English-to-Korean translations, including excessive passive constructions, unnecessary plural markers, and redundant subject pronoun repetitions, replacing them with natural native active phrasing.
• Breaking Rigid Parallelism: Generative LLMs routinely output monotonous compound sentence structures, predictable transitional adverbs, and uniform sentence lengths. The skill dynamically varies clause cadence, introducing natural linguistic flow and rhythmic variation.
• Zero-Loss Factual Integrity: Unlike broad paraphrasing utilities that inadvertently mutate key arguments, figures, or technical assertions during stylistic rewriting, the tool operates under a strict principle of zero-loss semantic fidelity, ensuring all primary assertions and underlying data remain intact.

Rather than performing haphazard synonym swapping, the engine restructures phrasing to match the natural breathing room of native Korean prose, transforming stiff outputs into compelling, publication-ready text.

Agent CLI Integration with Claude Code and Real-World Document Utility

Engineered to eliminate the friction of toggling back and forth across browser tabs with copy-paste prompts, 'im-not-ai' integrates straight into existing terminal agent workflows.

Released as an MIT-licensed open-source project, it is packaged as a ready-to-run skill tailored for next-generation developer agent harnesses including Claude Code and Codex CLI. Users can invoke document refinement commands directly inside their active command-line sessions alongside ongoing programming or research tasks.

• Thriving Open-Source Adoption: With over 5,000 GitHub stars, the project has gained substantial traction across the developer and writing community, serving as a trusted standard for post-processing AI-generated Korean text.
• Diverse Real-World Use Cases: The skill has seen enthusiastic uptake across a broad spectrum of practical documents, including job seekers refining cover letters and resumes, professionals polishing business proposals, students crafting academic reports, and engineers drafting technical documentation.
• Repository Context Cohesion: Because the agent operates directly within local project directories and version-controlled environments, it can align its revisions with established project nomenclature and style guidelines, maintaining consistent tone across large multi-file documentation repositories.

CLI Harness Setup and Key Considerations for Technical Review

Teams and practitioners evaluating 'im-not-ai' for their editorial toolchain should keep several operational constraints and best practices in mind.

First, because this tool is architected as an agent skill intended for CLI environments like Claude Code and Codex rather than a standalone desktop GUI or browser extension, users must have a functional CLI agent harness and compatible runtime properly configured on their local workstation.

Second, while stylistic smoothing significantly enhances readability, highly specialized domains involving specialized engineering terminology, legal covenants, or delicate factual nuances always carry potential risks of subtle semantic drift during automated transformations. Practitioners are strongly advised to keep a human in the loop, conducting a final review to verify that domain-specific technical vocabulary, numbers, and critical assertions remain completely accurate before formal publication.

Sources

• GitHub Official Repository: epoko77-ai/im-not-ai on GitHub — Official repository for the MIT-licensed Korean AI text humanizer skill (humanize-korean) supporting Claude Code and Codex CLI environments.
• 꼬마봇 Official X Announcement (@BBBang9900): @BBBang9900 on X — Practical workflow demonstration and community introduction to polishing machine-translated cadence in AI-drafted Korean text.