A Korean Writing Skill: Before and After
The rewritten drafts still sounded awkward
I had asked the AI to rewrite the blog build account, but its focus was still unclear and the Korean sounded awkward. Actual requests and verification records were present, yet the writing kept saying that it had connected requirements, implementation, and verification. An account of the work kept turning into an explanation of how the article had been organized.
I remembered seeing a Korean writing prompt or skill on GitHub a few months earlier and asked the AI to look for it. From the candidates, I chose to try hjongc/humanizer-kr first. I also asked for an article recording what changed during its use.
The AI revised two Korean posts and updated their English versions to match. This account records the edits and their reasons.
Giving each article a question before editing its sentences
Humanizer KR starts with the reader and the kind of text, then edits the wording while preserving facts and intent. It favors concrete actors and verbs over abstract nouns. If the user supplies a writing sample, that sample takes priority over generic style rules.
The AI read the skill instructions and used them to revise the drafts. The skill does not make every editorial decision. Its instruction to preserve the original claims means that an article with the wrong focus needs a separate structural edit. The AI divided this revision into two tasks:
- Reduce vague wording and repeated explanations of who did the work.
- Define the question each article answers and bring the actual requests and results forward.
For the build account, the question was: “What did I delegate to AI, and where did I have to redirect it?” For the revision account: “Why did a request to record actual work produce general advice?” The AI formulated these questions for this edit; they are not verbatim quotations from my earlier instructions.
What changed in the sentences
The Korean article quotes the preserved draft and the September 16, 2026 revision exactly. The examples below translate those Korean passages to show the edit; they are not verbatim quotations from the older English article.
A sentence about the folder structure
Before, translated:
The final structure takes a form that distinguishes the two roles under the same project.
After, translated:
Documents and executable code were kept separately under the same project.
The revised sentence names the two things being separated. It describes how they were kept instead of referring vaguely to a structural form.
A passage identifying who fixed a bug
Before, translated:
The entity that discovered and corrected this error was AI. My role was to request Korean and English support and to delegate verification of the actual screens and interactions.
After, translated:
The AI found the bug and changed the code. I requested bilingual support and delegated checks of the actual screens and interactions.
The requester and the person or system doing the work remain the same. A description of “my role” becomes a direct account of what I requested.
These are wording edits. Moving folder discovery, the rejected design, and the translation bug into the build article’s opening was a structural decision. The separate closing section that repeated the roles was also cut down. Those changes cannot be attributed solely to the skill.
Interpreting the automated findings
The AI ran the skill’s audit_korean_text.py on the original and revised files. Both scans used the same settings and covered all Markdown content, including metadata, code, links, and quotations.
| Findings labeled as nominalized wording | Before | After |
|---|---|---|
| Blog build account | 10 | 2 |
| Draft revision account | 16 | 6 |
These are counts of matches to the script’s rules, not writing-quality scores. Length and structure changed as well, so the decrease is not a controlled measure of the skill’s performance.
Slashes in code and links were also flagged as punctuation clutter. The AI did not change real quotations or code just to clear those findings. It read the remaining expressions in context. The script helped locate text to inspect; it did not decide the final wording.
The title revision
The titles needed another edit after the skill was applied. The original Korean title reported that I had tried editing the drafts. I objected to sentence-style titles and their lack of a hook. The replacement foregrounds the before-and-after comparison that readers can inspect. The English title is now “A Korean Writing Skill: Before and After.”
The other two titles now highlight design rework and a translation bug, and the real project details missing from the drafts. These changes followed additional feedback on the titles; they were not automatic suggestions from the skill.
Preserving the record while making the edit
Before editing, the AI saved the four Korean and English originals. It checked the quoted before-and-after passages against those files and the revisions. Quotations in the earlier revision account retained their wording and were labeled as excerpts from September 15. Historical quotations were not rewritten to match the latest article.
Publication boundaries, the choice of design 2, the translation bug’s cause, and the corrected code were preserved. The English versions follow the same facts and sequence. The Korean-specific skill was not used to validate English style.
The observable edits are narrower: repeated explanations were reduced, descriptions of roles became direct actions, and actual requests and results moved into the openings. No separate sample of my writing was supplied, so this does not establish that the AI reproduced my voice. Before publication, the two posts were reviewed again for wording and focus separately. Following the skill’s rules does not by itself establish that an article answers the original question.
The results are in the blog build account and the draft revision account.
The skill used was this revision of hjongc/humanizer-kr, published under the MIT license.
Applying the criteria to eight earlier cases
I later asked AI to turn eight topics selected from earlier work sessions into Korean and English articles. It returned to the linked original messages to check requests, corrections, and result reports instead of expanding the candidate summaries alone.
For example, a working title about investigating extra folders and preserving branches became “10 Working Folders Removed, 10 Branches Preserved”. Both counts came from the completion report, bringing the removal and preservation decisions into the title together.
The transcription article connects a 30-second sample’s processing time with repetition in the full recording. It does not treat the absence of that repeated pattern as proof of complete transcription accuracy. The Discord article similarly separates implementation and startup reports from the mobile workflow test still outstanding.
The new articles lead with specific problems rather than tool introductions. After editing the Korean text, AI checked that the English versions preserved the same actors, numbers, and unresolved checks. The observable changes are in titles, structure, and use of evidence. Reader comprehension and click-through rates have not been measured.
Follow-up comparison
A later comparison applied im-not-ai to the timer and session-topic articles. It separated a style-only pass from structural editing. Adopted and rejected changes are documented in the two-article editing account.