Rewrite an AI generated Korean draft so it matches a specific person's writing habits.
Build a personal style profile from a writer's own published posts.
Check whether a rewritten text still sounds like AI using automated style verification.
Adapt the same pipeline to clone a different writer's voice using their own corpus.
| gongnyang/gn-voice | albertcheng19/medskillos | amap-ml/roleagent | |
|---|---|---|---|
| Stars | 77 | 77 | 77 |
| Language | Python | Python | Python |
| Setup difficulty | moderate | moderate | hard |
| Complexity | 3/5 | 4/5 | 5/5 |
| Audience | writer | researcher | researcher |
Figures from each repo's GitHub metadata at analysis time.
Best results require your own sizable Korean writing corpus, not just the install script.
gn-voice is a Claude Code skill that rewrites AI generated Korean text so it reads like one specific person's own writing. The author built it because commercial editing tools smoothed their writing into a generic average that no longer felt like their own voice. The approach starts with data. The author collected 986 of their own published posts and ran them through Korean morphological analysis to measure real habits, such as which sentence endings they use, how often they use commas, and where they break lines mid sentence. These measurements are grouped into 23 cells by genre and channel, then distilled into short style packs, 9 based on the author's own writing plus 6 general genre packs such as speeches and sales copy, 15 in total. A decision tree picks which pack applies to a given piece of writing. The corpus itself is stored as plain text files, one file per post, with no database or search engine involved. Only writing the author wrote entirely themselves is included, AI assisted drafts are filtered out, and 15 percent of the collected posts are held back purely for testing so the system is checked against writing it has not seen. Rewriting happens through three separate roles instead of one. A runner picks the right style pack and rewrites the text in a single pass while keeping every fact and number unchanged. A gate then checks the result with plain Python scripts, comparing sentence endings, punctuation, and line breaks against the measured ranges and checking for common AI sounding phrases. A separate judge agent grades the final text from A to D against the style pack rules and points out specific lines that break them. The author is upfront about limits. AI detection models still flag most rewritten text as AI written, since the goal is to sound like the original author to human readers rather than to fool detectors. Genres with few sample posts get looser, direction only guidance rather than strict rules, and the author still expects a human to read the final output before publishing. Anyone can point the pipeline at their own writing by replacing the corpus folder, as long as the corpus contains only their own original work.
A Claude Code skill that rewrites AI drafted Korean text to match one author's real writing style, verified with measured corpus statistics rather than guesswork.
Mainly Python. The stack also includes Python, Claude Code, kiwipiepy.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly writer.
This repo across BitVibe Labs
Don't trust strangers blindly. Verify against the repo.