isaaceryn/humanizer-ko, released under the MIT License, is a Claude Code agent skill that finds and fixes traces of AI authorship in Korean text. In the commit b1562c26315b8d264d7830cd98d280e4d92333ea of October 21, 2026, the version was bumped to 1.5.0, and a new reference file, references/academic-papers.md, covering only paper abstracts and academic manuscripts was added. This commit adds 194 lines and deletes 8 lines, and the 37 pattern numbers and their counts in the body of the skill remain unchanged.
The starting point of this file is that the traces in an academic manuscript come from its structure rather than from its sentences. The patterns are divided into seven, A1~A7: A3~A7 are generation traces measured from text written entirely by an AI from the start, while A1 (abstract overstuffing) and A2 (clue accumulation) are treated as defects that arise when a human-written manuscript is polished with AI several times over. It also contains a measurement result whose direction runs counter to expectation: AI-written abstracts had fewer results and fewer numbers than human-written abstracts, and longer conclusions. So a text is not judged to be AI-written merely because it contains many figures; judgment is made by a cluster of patterns rather than by a single pattern. What to keep when shortening an abstract within a limit (problem, method, key results, implications) is written out in order in the file, and the English abstract vocabulary list is annotated with the measurement basis and the supporting literature. SKILL.md also gained one more evaluation guideline: do not rewrite into AI-style forward-looking promises either the guidance on section structure at the end of the introduction or the list of research questions.
That said, the weight of the evidence is not even. The only thing the file itself measured directly on Korean abstracts is the comma after a connective ending (§37), and it explicitly states that the remaining generation traces are hypotheses carried over from English measurements. There is no measurement study for abstract overstuffing or clue accumulation, so judgment has to rely on the abstract guidelines of academic journals. In other words, 1.5.0 did not improve the detector; it expanded the polishing rules in the form of documentation, and its accuracy for Korean academic writing has not yet been demonstrated by the project itself. Since the rules take a form a human can read and apply as is, they are worth trying when shortening an abstract or checking a submission.