Surgically Editing Korean AI Drafts

Definition

A surgical Korean-editing workflow detects language-specific AI-writing patterns, rewrites only the affected spans and then checks that the edit did not drift too far from the original. It differs from asking an LLM to rewrite the whole draft because meaning, genre and the writer's existing voice are explicit constraints.

Perspectives

lucas (2026-08-22, X)

Prefer rule-based, span-level correction over a full LLM rewrite when reviewing Korean AI-assisted writing. Look for translationese, mechanical parallel structure, passive-voice overuse, sentence-initial conjunctions and stock conclusion phrases, then change only those parts. This preserves the original writer's style and intent while improving Korean naturalness, and is easy to attach to a working writing pipeline.

Humanize KR maintainers (2026-08, GitHub)

Preserve facts, claims, numbers, proper nouns and direct quotations; edit only detected spans; retain the source genre; and stop over-editing with a 30% change warning and a 50% hard limit. Route a clean draft to one conservative call, a typical AI draft to diagnosis plus one targeted edit, and severe, evidence-sensitive or over-15,000-character material to diagnosis, editing and final verification. Prefer one edit call below 15,000 characters because repeated rulebook loading can make chunked editing much more expensive without improving quality.

How to apply

• Best suited to a completed Korean draft whose meaning and voice are already acceptable but whose surface language still feels machine-written.
• Use Humanize KR when you want the detection taxonomy, severity routing and deterministic change-rate gate packaged together.
• For generation-time Korean instructions rather than post-draft correction, compare Fluent Korean.
• Apply after drafting or repurposing with an AI content skill graph, not before the core idea and channel format are settled.

Limits

• The workflow, thresholds and token figures come from the repository and one practitioner's recommendation; no independent quality evaluation was verified.

• Naturalness detectors and pattern taxonomies can mistake deliberate rhetorical style for an AI artifact, so a person should review high-change outputs.

• Multi-call routing is described as available only in Claude Code; Codex, Copilot and Gemini use a single-call path in the captured version.

• Original X post