Humanize Korean: Strip the AI Tell from Korean Text

Where this fits

You run a Korean-language content pipeline where every draft — blog posts, product copy, translated documentation — comes out of an LLM with the same mechanical rhythm: heavy passive voice, uniform sentence length, bullet-point tics, and English-translation residue. Your editors spend more time de-AI-ing text than editing it. This skill sits in that gap: it takes an AI-generated Korean draft and rewrites only the style layer — rhythm, idiom, voice — while leaving every fact and claim intact. It is a post-processing step, not a content generator.

Why agents benefit

• Detects and classifies 40+ AI tell patterns across 10 categories (translation-ese, excessive English quotes, mechanical parallelism, passive-voice abuse, connector overuse, uniform rhythm, emoji/bullet overload) in a single pass.
• Rewrites style and rhythm only — content is untouched, so you can chain it after generation without risking factual drift.
• Responds to follow-up commands: "only this category again," "adjust intensity," "change genre," "just this paragraph," "second pass" — all handled by the same skill.
• Offers two execution modes: Fast (monolith, 4-5 tool calls, 2-3 min for 5,000 chars) and Strict (5-agent pipeline for long documents and precision audits).
• Draws clear boundaries: spell-checking stays local, translation goes to a translation skill, and content-level rewrites are routed elsewhere — so the agent never oversteps.

Practical scenarios

Content marketing localization

A Korean SaaS blog imports English posts, machine-translates them, then needs them to read like a native marketer wrote them. The skill strips translation-ese and English-quote residue in one pass, cutting editorial review time roughly in half.

Academic and submission writing

Students and researchers draft papers in ChatGPT, then need to remove the telltale AI cadence before submission. Strict mode's auditor pass catches patterns the fast path misses — useful for longer documents where detection precision matters more than speed.

Multi-round editorial refinement

An editor runs a draft through fast mode, then asks for "only the connector overuse category, less aggressive" on the second half. The skill's category-specific follow-up commands make it a reusable editorial tool rather than a one-shot transformer.

Add it to your agent workflow

Install via the skills CLI:

npx skills add bam-bam-2/solo-skills --skill humanize-korean

Then trigger it in your agent with natural language:

User: 이 블로그 초안 AI 티 없애줘 (strip the AI tell from this blog draft)
Agent: Runs humanize-monolith in fast mode, returns rewritten draft + pattern report

For stricter handling of a long document:

User: --strict 모드로 이 논문 초안 윤문해줘
Agent: Runs the 5-agent pipeline with full audit

Compare before adopting

Before committing, compare against three things. First, quality signals: does it actually reduce AI-tell patterns, or just paraphrase? The v1.x changelog shows honest iteration — v1.2–v1.4 were rolled back for slow hot-path costs, which is a good sign of disciplined engineering. Second, alternatives: generic rewriting prompts or other humanizer tools may be cheaper but lack the 10-category taxonomy and follow-up controls. Third, workflow fit: if your pipeline is pure Korean content that needs style-only post-processing, this fits tightly; if you need content-level rewriting or translation, it explicitly defers to other skills.

Why it is worth tracking

The skill has 128 stars, an active version history, and a clearly documented boundary between what it does and what it hands off. Its fast/strict mode split addresses a real operational concern — latency vs. precision — that most humanizer tools ignore. Track it if you build Korean-language agent workflows where output authenticity matters, and evaluate it against your own AI-tell test set before wiring it into production. You can find the full skill page and community context at OpenAgentSkill.