humanize-text

Imported from korECM/humanize (SKILL.md). Install upstream with npx skills add korECM/humanize. Copyright stays with the author.

Humanize Text

Rewrite text so it reads as human-written. Language-agnostic procedure + per-language rule packs.

Procedure

1. Detect language of the input.

   • Use scripts/detect-lang.js <file> if available, or judge from script + vocabulary.
   • Supported: ko, en. Fallback to closest match for other languages, but warn the user.

2. Load the rule pack for that language: references/<lang>.md.

   • Always also load references/_common.md for language-agnostic principles.

3. Diagnose AI tells in the input against the rule pack's ## AI tells section.

   • List the specific tells you found (line/phrase level). Do not skip this.
   • Also note the human traits already present (voice markers, quirks, distinctive phrasing) — these must survive the rewrite. See _common.md ## Minimum effective edit.

4. Rewrite applying the rule pack's ## Rewrite rules. Then apply ## Human signals to inject naturalness.

   • Preserve meaning, facts, numbers, names, structure of argument.
   • Do not invent new claims, opinions, or anecdotes the original does not imply.
   • Edit only diagnosed spans; leave sentences that already read human untouched.

5. Validate against _common.md metrics:

   • Burstiness — sentence-length variance (mix short and long).
   • Register variation — not every sentence in the same form.
   • Concrete > abstract — abstract nouns reduced.
   • No meta scaffolding ("In conclusion / 결론적으로 / 따라서") unless the original explicitly needs it.
   • Change volume proportional to the tells found — if most of the text changed, re-check for meaning drift and over-polish.

6. Optional deterministic post-pass: if the working tree has scripts/rules/<lang>.json, you may run the regex substitutions in it as a final pass. These are safe one-to-one swaps (translationese phrases, AI vocab blocklist).

Output format

Default: return only the rewritten text. No preamble, no explanation.

If the user asks "what did you change" or "diff", produce a brief diagnosis list (the AI tells found, the rules applied) followed by the rewritten text.

Detect-only mode

If the user asks whether the text reads as AI-written, or to audit/scan/flag a draft without rewriting: run steps 1–3 only. For each tell found, quote the phrase and give the fix in a few words. Do not rewrite, do not produce a score, and do not claim to know who wrote it — detectors guess, named patterns are evidence the user can check. Offer to rewrite afterward.

When NOT to humanize

• Legal, regulatory, academic-formal contexts where the formal register is required — confirm with the user first.
• Code, command output, structured data (JSON, YAML).
• Direct quotations.

Adding a new language

1. Create references/<lang>.md with the same section structure as ko.md: ## AI tells → ## Human signals → ## Rewrite rules → ## Examples
2. Create scripts/rules/<lang>.json with { "translationese": [[pattern, replacement], ...], "blocklist": [...] }.
3. Add the language to scripts/detect-lang.js's supported list.

The procedure above is language-agnostic. Rules live entirely in the language pack.