A Korean-language humanizer that detects and rewrites AI-sounding text (translationese, mechanical parallelism) while preserving facts, with deterministic gates.
A well-engineered, narrowly scoped Korean editor with route selection and verification scripts; useful only if you write Korean.
Humanize KR rewrites Korean text that reads as machine-written without changing its facts, figures, names or quotes. It does not detect AI text for a score; it edits. The README classifies patterns into 10 categories and 85 sub-patterns graded by severity, such as English-style translationese, mechanical enumeration (first, second, third), stock closers like "it has significant implications", overused passives, sentence-initial conjunctions, and emoji or bullet overuse.
A Python script first scores the input and picks one of three routes: light (one model call for already good text), standard (diagnosis plus a targeted rewrite), or heavy (diagnosis, rewrite, then a finalizer that checks meaning preservation, used for very bad or very long text). A gate script then checks change rate (warning above 30%, stop above 50%), numbers and quoted text. Four rules are fixed: meaning unchanged, edits only on detected spans, genre kept, no over-editing. Pasted text is treated as data, not instructions.
The main skill is about 31 KB with 18 reference files and two metric scripts; Codex and Copilot versions are provided separately, along with a commit-ko skill that cleans up Korean commit messages. The README lists Claude Code, Codex, Copilot CLI and Gemini CLI, and says the Gemini extension uses a smaller, separate rulebook.
Seven skills, the main one 31 KB with 18 reference files and metric scripts, plus Codex and Copilot variants and gate scripts, but no evals.
Counted from the repository file tree. How we score depth.
A Korean text-humanizing skill with local Python checkers. The installer creates symlinks in your Claude, Codex and Gemini skill and agent folders, and update.sh can pull new commits from the upstream git remote and re-run the installer when you run it. No network calls or credential access were found in the scripts.
echo Y | gemini extensions link <repo>, answering that CLI's confirmation prompt automatically; update.sh runs git fetch and git pull --ff-only on the checkout and re-runs install.sh when you invoke it.install.sh:231The repository (159 files) on Oct 29, 2026. Not a full review: we read 5 of 69 code files; it relies on code outside the repository: Not read in full: the Python scripts under scripts/, skills/humanize-korean/references/metrics*.py and the test suite (inspected by grep for network, subprocess and file access); gemini CLI invoked by install.sh; community opencode port at im-not-ai-ocx.illuwa.click linked from the README (separate project).
NVIDIA SkillSpector static scan: 50 pattern hits (Dangerous Code Execution 17, Agent Snooping 16, Tool Misuse 8, Data Exfiltration 8, Supply Chain 1). HIGH hits are test fixtures copying os.environ for subprocess runs, rm -rf of temporary HOME directories in install tests, a documented DELETE route in a web-service spec, and the repo's own .githooks/commit-msg. 'Skill Enumeration' and 'File System Enumeration' hits are README and INSTALL commands such as ls ~/.claude/skills. The README's voice_profile_log.json telemetry is a local log file; no upload code was found.
A review, not a certification. Skills change after we check them — read the files before installing.
MIT — commercial use allowed.
Not really. It runs per text with deterministic gates, but it is an editing aid you review rather than a long unattended loop.
A lightweight CN/EN academic-writing skill that restructures paper drafts around one strongest claim instead of hedged, evenly-weighted self-critique.
13 skills for writing Chinese web novels: ranking scans, teardown of hit books, long and short drafting with file-based continuity tracking, de-AI editing and covers.
About 65 skills that apply frameworks from well-known business, marketing, UX and software books, plus guided multi-step journeys that keep state in your docs folder.
It is an agent skill that rewrites Korean text written by ChatGPT, Claude or Gemini so it reads like a person wrote it, by fixing translationese, mechanical parallelism and stock phrases. It keeps facts, numbers and quotes unchanged.
In Claude Code run /plugin marketplace add epoko77-ai/im-not-ai then /plugin install humanize-korean@im-not-ai. Copilot CLI, Codex (via install.sh) and Gemini CLI (gemini extensions install) are also covered in the README.
No. The patterns are Korean-specific. For English text use a general humanizer skill instead.
We saw only local Python scripts and instructions, and no network calls in the code we read. The installer symlinks the skills into your agent folders.