blog-humanize-fast

Blog Humanize Fast

the blog final manuscript meaning-preserving method. This skill is not a drafting skill. From the Korean blog body text already written by the author, it conservatively reduces AI tells such as translationese, mechanical parallelism, idioms, overuse of conjunctions, and uniform rhythm.

Input

• File path: .md, .mdx, .txt body file.
• Pasted text: 8,000 characters or fewer recommended.
• For MDX with frontmatter, if possible create a temporary body file excluding the frontmatter and run it. Frontmatter field values are not subject to polishing.

Execution

1. In this skill directory, references/quick-rules.mdis used as the rulebook.
2. If the Workflow tool is available, run workflows/blog-humanize-fast.js.
   • workflowPath: {SKILL_DIR}/workflows/blog-humanize-fast.js
   • args: { inputPath, genre: "블로그", intensity: "보수", quickRulesPath: "{SKILL_DIR}/references/quick-rules.md" }
3. If the Workflow tool is not available (including other runtimes such as Codex), this skill cannot be executed. Explicitly tell the user to run it in Claude Code, and if the current runtime has humanize-korean skill, guide them to its Fast Path as an alternative. Do not imitate polishing with an ordinary chat response.

Behavior contract

• Meaning invariance is the top-level rule. Facts, claims, numbers, dates, amounts, proper nouns, direct quotations, code, URLs, Markdown structure, and JSX tags/attributes are not modified.
• YAML frontmatter, code fences (```·~~~), and indented code blocks indented by 4 or more spaces are isolated as protected chunks and preserved byte-identical. Original newlines between chunks are left as is, so reassembly must be identical to the original when unchanged (integrity gate).
• If per-chunk fidelity verification fails, retry only once, and if it fails again, roll back that chunk to the original text. If a chunk's change rate exceeds 50%, roll it back regardless of fidelity, and if the overall change rate exceeds 30%, leave a warning.
• Global naturalness and residual S1/S2 detection are performed as separate axes.
• Targeted fixes are performed for only 1 round on only the pointed-out spans, and the revised version is adopted only when it passes the fidelity gate (if it fails, keep the body before the fix).
• Results are not applied automatically. Report the final version, change statistics, and residual count, and apply to the original file only when the user accepts.

Fallback

• Report as-is the causes of Workflow execution failure, quick-rules read failure, and model subcall failure.
• If the user requests more precise verification or the result grade is low, guide them to re-verify with humanize-korean --strict.