Humanizer-zh: a Chinese-language editing skill for Claude Code

The problem Humanizer-zh addresses in Chinese drafts

Chinese text produced by a language model tends to fail in recognizable ways: stacked four-character phrases, "进行+verb" constructions, passive chains, decorative headings, and closing lines that promise significance without stating anything. Humanizer-zh is a Claude Code skill that targets those patterns in existing articles, comments and documents. The README frames the job as editing, not detection: the input is a passage or a file, the default output is a finished rewrite, and sentences with no problem may be left untouched. The audience is narrow and specific. You need Claude Code, you need to be working in Chinese, and you need an editor that keeps the author's stance rather than flattening everything into the same conversational register. The README is explicit that this is an editing guide executed by an agent, not a standalone detector, and that it cannot prove who wrote a text or guarantee it passes any AI detector. That disclaimer is the most important sentence in the repository, because it rules out the use case most people arrive with.

How the skill works: SKILL.md, 31 checkpoints and a priority order

There is no runtime, no model and no scoring script. The repository holds SKILL.md as the skill definition, README.md as documentation, a CHANGELOG.md, tests/, and an MIT licence. When Claude Code loads the skill, the agent reads SKILL.md and applies its rules to whatever text the user supplies. The rules are organized as 31 checkpoints in six categories carried over from PR #39: A covers setup that replaces statement (fake contrast, dramatic fragments, pseudo-depth, run-up openings, defence with no object); B covers formulaic rhythm (forced triads, repeated openings, the universal em dash, stacked qualifiers, invented compounds, passive voice with a missing subject); C covers inflation and borrowed authority (empty high-frequency words, significance inflation, vague association, sentence-final uplift, promotional language, authority endorsement, complex copulas); D covers formulaic layout (useless bold, decorative headings, quotation marks and punctuation); E covers chat and draft residue (customer-service tone, repeated disclaimers and speculative padding, first sentence restating the title, editing-process leftovers); F adds Chinese-specific checks (long attributive phrases, "进行+verb", passive stacking, four-character parallelism, universal background, formulaic closings). The priority order matters more than the list. The README states the constraint order as: preserve information and certainty, respect the user's scope and register, match the author's voice, and only then fix expression problems. A pattern match cannot override those. Concretely, "可能" must not become "确定", "计划" must not become "已经", and negation, conditions, attribution, scope and completion status survive the rewrite.

Installing Humanizer-zh and running it on one file

The README gives three installation routes. The recommended one uses the skills CLI, which places the skill in the correct directory automatically.

npx skills add https://github.com/op7418/Humanizer-zh.git

The second route is a git clone into the Claude Code skills directory, which is the path to use if you want to read SKILL.md before it is loaded.

git clone https://github.com/op7418/Humanizer-zh.git ~/.claude/skills/humanizer-zh

The third route is manual: download or clone the repository, copy the Humanizer-zh folder into ~/.claude/skills/ on macOS and Linux, or %USERPROFILE%\.claude\skills\ on Windows, and confirm the folder contains SKILL.md and README.md. After restarting Claude Code or reloading skills, the README says typing the command below activates the skill if the install worked.

/humanizer-zh

For a first real use, the README's file-editing example is the safest starting point because it keeps the scope bounded. It also documents a review-only mode: state that you want suggestions and no file changes. If you have a writing sample from the author, you can supply it; the README says the skill borrows expression habits from the sample but will not move the sample's experiences, data or opinions into the target text.

What the rewrite examples actually demonstrate

The README ships four worked examples and labels them teaching inputs with no hidden supplementary material, which is the honest way to present them. The first is a product blurb reading "该方案稳定可靠、快速响应、易于维护" rewritten as "这套方案运行稳定、响应快，也方便维护". Nothing was added: no launch date, no response-time figure, no maintenance cost. The README's point is that missing evidence is a reason to ask the author, never a reason to invent a number. The second example handles uncertainty. A bug report saying a problem is "被认为可能与内存泄漏有关，但原因尚未确认" keeps both the hedge and the open status, and the rewrite does not hand the causal judgement to the community. The third example is the one that shows the rules are not mechanical: a release note with three genuine features keeps all three, because avoiding a triad is not a reason to delete a real item or pad it. The fourth example is a two-sentence instruction that comes back unchanged, which establishes that the skill is not required to produce an edit on every call. Read together, these examples describe an editor whose default failure mode is under-editing, and the README treats that as correct behaviour.

Where Humanizer-zh is the wrong tool

The README's own disclaimer is the first limitation: this is an editing guide executed by an agent, it cannot prove authorship, and it does not guarantee passing any AI detector. If your requirement is a verdict about whether a text was machine-written, this project does not offer one, and the 31 checkpoints are explicitly described as an editing checklist rather than a standard for detecting authors. The second limitation concerns evidence. The verification section describes 18 short-text cases, one Markdown file sample and a structure-check script, plus a long-document comparison between the old and revised versions checking numbers, conditions, attribution, author attitude and file structure. The README then states that a single run on limited samples cannot represent all models or registers, and that a drop in word count or a model's self-score cannot prove effectiveness. That is an unusually direct admission that the project has no quantitative evidence for its own output quality. The third limitation is structural: the skill runs inside Claude Code, so it inherits that tool's model behaviour, and the README notes that file editing preserves code, commands, paths, link targets, YAML, data, headings and anchors by default, with the warning that changing a heading or structure requires checking references. If you work outside Claude Code, or you need a batch pipeline that processes thousands of files without an interactive agent, this repository gives you a rule set to port, not a program to run.

How it differs from blader/humanizer and stop-slop

Humanizer-zh is a localization and extension of blader/humanizer v3.0.0, and the README credits that project as the source of the A through E categories, the voice-calibration idea and the file mode. The difference is category F. The six Chinese-specific checkpoints cover long attributive phrases, "进行+verb", passive stacking, four-character parallelism, universal background passages and formulaic closings, none of which the English original needs to name. A second difference is the revision history: PR #39 supplies the structural basis and the Chinese checkpoints, and PR #34 is cited as a reference for example fidelity, with the README noting that this revision did not merge that branch directly. If you write English, blader/humanizer is the closer fit and the Chinese checkpoints will not fire. hardikpandya/stop-slop is credited as a reference for concise expression and editing checks, which points at a different emphasis: brevity and cutting, rather than preserving hedging and attribution. The Wikipedia page "Signs of AI writing" is named as the observation source behind the original project, which is worth knowing because it means the checkpoint list descends from an encyclopedia's maintenance heuristics, not from a linguistic study.

Maintenance, licence and what an edit costs you

The repository is not archived and the last push was on 2026-09-23, five days before this writing, with that same date marked in the README as a rewrite of the rules and examples. There are no retrieved releases, so there is no versioned artefact to pin; you track the main branch, and the CHANGELOG.md is where the 2026-09-23 change is documented. That matters for upgrade cost. Because the skill is a Markdown instruction file rather than a package, updating means pulling the repository again or reinstalling through the skills CLI, and any local edits you made to SKILL.md will conflict. The practical approach is to keep your own additions in a separate note rather than inside SKILL.md. The licence is MIT, per the LICENSE file the README points to, which permits commercial use and modification provided the copyright notice and permission notice are retained; the README also lists the upstream projects it draws on, and if you redistribute a modified version, those attributions are the part to keep intact. This is a description of the licence text, not legal advice. The per-edit cost is the real budget item: every pass consumes agent context, and the README's default of delivering only the final draft means you do not get a hit list or a self-score to review, so quality control falls back on reading the output yourself.

Editorial conclusion

Adopt Humanizer-zh if you write Chinese prose in Claude Code and want a checklist-driven editor that refuses to invent facts. Do not adopt it if you need proof about authorship or a pass on an AI detector; the README says it cannot provide either. Before trusting it, read SKILL.md's 31 checkpoints and tests/README.md, then run it on one file with the instruction to give suggestions only and change nothing.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. op7418/Humanizer-zh on GitHub
4. README