Korean Tech Writing as a Pipeline, Not Taste

Where this fits

A developer finishes a feature and needs a Korean tech-blog post for the company engineering blog. They draft it, but the prose reads like translated English — passive, indirect, with floating numbers like "5x faster" and no clear audience. ko-tech-writer turns that cleanup into a fixed pipeline: register detection, real-sample calibration of the target venue, translationese removal, honesty-level layering, a number/claim gate, and a rendition plus perceptual QA pass. The guiding principle: good sentences are a procedure, not a matter of taste.

Why agents benefit

• Venue calibration from real samples: The agent fetches actual published posts from the target blog to calibrate tone, instead of guessing what "blog style" means.
• Translationese removal as a pattern list: It applies a concrete checklist of Korean translationese patterns (passive overuse, "-에 의해" chains, awkward noun stacking) rather than a vague "make it sound natural."
• Honesty-level layering: Reader-facing claims and internal audit vocabulary are kept separate, so the published post never overstates what internal metrics actually show.
• Number/claim gate: Every figure must carry a source and scope window. The agent rewrites or flags "3x improvement" until it is backed by a baseline, a tool, and a measurement date.
• Perceptual QA: The pipeline ends by reading the rendered output for rhythm and formatting issues — catching problems a pure LLM review often misses.

Practical scenarios

Revising a draft full of translationese

A developer pastes a draft that was written in English and machine-translated. Trigger: "이 문서 번역투 고쳐줘". The agent detects the register, pulls two or three real posts from the company blog for tone calibration, and rewrites using the translationese pattern list.

Producing a talk script and teaser from one post

The same engineer is presenting at a meetup. The agent takes the approved blog post and generates a spoken script (shorter sentences, conversational rhythm) plus a teaser summary — each with its own register and compression rules.

Enforcing number hygiene

The draft says "성능이 3배 개선". The number/claim gate asks: from which baseline, measured with what tool, over what window? The agent proposes "이전 버전 대비 요청당 지연 시간 38% 단축 (benchmark X, 2026-07 측정)" or flags the claim for the author to source.

Add it to your agent workflow

npx skills add chrono-meta/forge-harness --skill ko-tech-writer

Then invoke it with one of the Korean triggers:

"이 글 회사 기술블로그에 올릴 수 있게 다듬어줘"

Compare before adopting

• Quality signals: Run it on a document you have already hand-edited. Does the pipeline match or beat your edits? Compare translationese removal rate and whether the number gate catches real gaps.
• Maintenance freshness: The rule classes come from an actual 2026-08-10 editing session. Check the commit history to see if the pattern list is updated as new classes are detected.
• Alternatives: Generic "Korean proofreading" prompts fix grammar but rarely calibrate to a venue or gate claims.
• Workflow fit: The skill expects tools like Read, Write, Edit, Bash, Grep, Glob, and WebFetch. Confirm your agent runtime allows file access and web fetching for sample calibration.

Why it is worth tracking

ko-tech-writer is worth tracking because it treats editorial quality as an auditable pipeline, not a personality trait. Real-sample calibration and the number/claim gate are the differentiators — they map directly to real publishing failures like overstated metrics, translationese, and tone mismatch. At 14 stars it is early, so evaluate it on your own documents before depending on it. More context and related skills are listed at https://www.openagentskill.com/skills/chrono-meta-ko-tech-writer.