Editing Two Blog Posts: Clearer Sentences, Less Repetition
Before adding another writing skill
The recent posts still repeated explanations after an earlier Korean editing pass. Conclusions restated my requests and AI’s work, and paragraphs about organizing the article took up space alongside the actual case.
The skill used so far was hjongc/humanizer-kr. When I asked what to try next, AI suggested comparing two posts with im-not-ai, the first candidate in the earlier research, and editing their structure separately. I asked it to proceed and write about the work too.
The subjects were the speech timer and the selection of topics from previous sessions: an app-building case and an account of research and editing.
Three versions from the same starting text
AI saved the Korean and English originals, then produced this comparison for each Korean article.
| Version | Scope of change |
|---|---|
| Current text | The existing draft, already edited with humanizer-kr |
| Sentences revised | Only expressions covered by im-not-ai; facts and paragraph order retained |
| Order and repetition revised | The same starting draft, with separate changes to the opening, section order, and repetitive conclusion |
The structural versions also incorporated useful changes from the style pass. These were not outputs from three different models. The same AI used different instructions and editing scopes; there was no reader-preference test.
Installation needed a repair too. The Codex skill’s reference directory arrived as a file containing a link path rather than the referenced documents. AI retrieved the documents from the same pinned version before reading the rules. The installation message alone had not established readiness to apply them.
The style pass made small changes
The timer’s first Korean heading changed from a construction equivalent to “What the speaker needed to see was the background color” to “The speaker needed to see the background color.” It preserved the requirement in fewer words. That change was retained in the structural version.
A longer sentence was split as well. The following is a translation of the comparison in the saved Korean drafts, not an English output from the skill.
Before:
AI reported reproducing the error in the browser and finding a typo in the code that handled the button list.
Sentences revised:
AI reported reproducing the error in the browser and finding a typo in the code. That code handled the button list.
The information remained, but the explanation now used two sentences. The final structural edit did not adopt that wording as written. It established the historical report as the source, then described the cause and checks. A sentence-splitting rule did not settle which version read best.
In the topic-selection article, AI removed the general announcement that selected candidates also had outstanding checks. The next sentence already specified the missing Discord test. It also removed commas after Korean sentence endings such as those meaning “and” while preserving punctuation in direct quotations.
The skill’s change-rate calculation returned:
| Style-pass draft | Change rate from its source |
|---|---|
| Timer | 1.52% |
| Topic selection | 0.48% |
The comparison covered Markdown body text, including headings, but excluded frontmatter and the check summary. It measured character-level differences, not improvement in writing quality. Both sources had already received an editing pass, so these figures cannot rank the two skills.
Repeated conclusions and openings needed separate edits
The style-pass timer article still ended with a section restating three corrections: removing numbers from the speaker view, clarifying temporary input, and requesting a log check after an error. Each had already appeared in its own section.
The structural version removed that conclusion. Requests and AI changes remained in the relevant scenes, and the article ended with the historical report of the Safari check from input through saved record. It retained the source note and the limit on unverified long-term use. The section count fell from four to three.
The topic-selection article opened with a general introduction to the task. Its revised opening starts with the timer display correction and the study request that produced a subtitle file, followed by the original request to search previous sessions. Readers encounter examples of the selected material immediately.
Deferred topics moved next to the selection criteria. The outstanding Discord test moved into the source-checking section. The repeated explanation of roles at the end was merged into the opening. Six sections became five. These were separate editorial decisions, not automatic effects of the style skill.
What was adopted and what was retained
Both structural edits were adopted, with English versions revised to match their content and order. The style-pass drafts remain comparison material. The Korean skill was not used to edit English; the translations were checked for the same actors, numbers, and unresolved results.
Actual quotations, the transcription-accuracy limit, the outstanding Discord test, and attribution to historical AI reports were preserved. Removing those details for brevity could change what the articles claim.
The style pass made small changes in this comparison. The openings and conclusions changed through separate editing. The resulting workflow checks expressions with a skill, then separately asks whether information is repeated and whether the opening presents the actual problem.
The instructions came from this pinned revision of im-not-ai. This was a two-article comparison on September 16, 2026; improved author satisfaction or reader response has not been established.