op7418/Humanizer-zh, a Chinese Claude Code Skill that strips AI patterns from text, hit GitHub daily trending with 18,433 stars and 249 added in one day S¹. The tool's own README admits it cannot prove who wrote an article or guarantee it will pass any AI detector S¹.

My read: This is the first Chinese-localized Claude Code Skill I've seen gain this kind of traction. The parent project, blader/humanizer, already sits at 38,309 stars P⁴, so the appetite for AI-text-scrubbing tools is clearly global. What I find honest is the maintainer's explicit disclaimer that the tool won't fool detectors, because it frames the skill as an editing aid rather than a deception engine. Whether 31 checkpoints actually make Chinese AI text read more naturally is something nobody outside the repo has independently tested.

The skill is not a standalone detection program. According to the maintainers, it is an editing guideline that a Claude Code agent reads and executes, targeting empty rhetoric, repetition, and templated phrasing in existing Chinese articles, comments, and documents while preserving facts, the author's certainty level, and their voice S¹. The project traces its core files to blader/humanizer v3.0.0, which itself draws on Wikipedia's "Signs of AI writing" guide [S1, P5], and its utility components to hardikpandya/stop-slop P⁵.

A 2026-09-23 update note in the README describes a full rewrite of the rules and examples to clarify how the tool preserves facts, author stance, and writing style S¹. The skill now carries 31 editing checkpoints organized under an A through F classification system inherited from Humanizer-zh pull request #39 S¹.

What the 31 checkpoints actually do

The editing principles prohibit fabricating functions, data, sources, identities, or first-person experiences S¹. The skill is meant to clean up how something is written, not invent what it says. Categories A through E come from the original English humanizer by the developer blader, while the Chinese-specific checkpoints were built on top of that structure S¹.

The project is released under an MIT License and was created on 2026-01-19 P², one day after the original blader/humanizer repository appeared on 2026-01-18 P⁴. A deep-research snapshot shows 16,252 stars and 1,088 forks at capture time P², while the trending listing recorded 18,433 stars S¹. That is a gain of more than 2,000 stars in the interim.

Claude Code Skills, the extension format this tool uses, have been accumulating momentum across the ecosystem. As we found when Claude Code v2.1.239 added cost tracking and hit 142k stars, the platform's star count has been climbing steadily since August. The swarm tool Ruflo reached 69,000 stars around the same time, as we noted in our coverage of Ruflo's GitHub traction, and a system for spawning Claude Code agents with pre-loaded memory appeared days later.

Who would use this, and what to check

A content team publishing Chinese-language technical documentation or marketing copy through Claude Code could load this skill to have the agent revise drafts for natural phrasing before publication. The practical check: run a sample paragraph through the skill, then compare the output against the original for any introduced factual claims. The maintainers' own rule against fabrication means any new fact in the revised text is a red flag.

All claims about the tool's effectiveness come from the maintainers' README. No independent review or benchmark has been published. The tool is not affiliated with Anthropic or Claude Code as an official product S¹. The 2026-09-23 update date appears in the source as written.

The repository currently shows 27 open issues P². Whether those issues surface community-reported edge cases in the checkpoint system, or whether the star count continues past 20,000, is the next thing to watch.

Sources: S1 — op7418/Humanizer-zh: Humanizer 的汉化版本，Claude Code Skills，旨在消除文本中 AI 生成的 · P2 — op7418/Humanizer-zh · P3 — ndpvt-web/arxiv-claude-skills · P4 — blader/humanizer · P5 — GitHub - op7418/Humanizer-zh: Humanizer 的汉化版本，Claude Code Skills，旨在消除文

Written from 5 sourced items, 4 of them primary.

GitHub stars: Humanizer-zh vs parent project blader/humanizer