Humanizer Writing Skill
Detects and rewrites common AI-writing signals so text sounds more natural and closer to a real author’s voice.
The skill declares only Read, Write, Edit, Grep, Glob, and AskUserQuestion, with no network, credential, or external-effect capability shown. However, Write/Edit access is broad and there is no explicit confirmation flow, sensitive-data handling, data-flow disclosure, or rollback procedure. The MIT license is clear, but publisher identity is unverified, so points are deducted.
The instructions contain extensive rules, examples, and semantic-preservation requirements, making the happy path plausible. There are no tests, CI evidence, abnormal-input procedures, or clear failure messages, and the referenced Process and Output section is not fully present in the supplied evidence. The score is therefore limited by static calibration and further reduced.
The audience, core input task, and optional voice-sample workflow are reasonably clear, and compatibility is declared as any-agent. Boundaries for false triggering and non-fit use cases, output formatting, and Chinese-language behavior are not sufficiently specified, so points are deducted.
The materials provide a name, version, installation instructions, usage examples, 33 pattern categories, version history, source attribution, and an MIT license. Documentation is readable, but runtime assumptions, troubleshooting, maintenance ownership, and update responsibility are incomplete, with minor detail inconsistencies between README.md and SKILL.md.
The broad rule set and many before/after examples plausibly support the core task of removing common AI-writing signals. There is no committed output evaluation, test evidence, or comparative evidence, and rigid rewriting, paragraph-preservation, and personality guidance may alter meaning or author voice. The output format is also unspecified, so points are deducted.
The README identifies Wikipedia and WikiProject AI Cleanup as methodological sources and includes a version history, providing limited auditability. There is no CI, test suite, reproducible experiment, or third-party execution evidence; conclusions therefore rely mainly on static claims and are constrained by the static-review cap.
- Write/Edit permissions may rewrite user text directly; confirm scope and retain the original or a recoverable copy.
- The rewriting objective may conflict with technical, legal, academic, or factual-accuracy requirements; review meaning, citations, and author voice manually.
- No Chinese-specific rules or execution evidence are provided, so Chinese-language effectiveness and false-edit risk are uncertain.
- The methodology relies on a Wikipedia guide, and the publisher is not verified by the curated enterprise registry.
What it does & when to use it
Humanizer is a portable agent skill for reducing recognizable signs of AI-generated writing. Based on Wikipedia’s “Signs of AI writing” guide, it covers 33 patterns across content, language, style, communication, filler, and hedging. It performs an initial rewrite, an audit for text that still looks obviously AI-generated, and a second rewrite. Users can also provide writing samples for voice calibration.
It reads the user’s text and, optionally, two or three paragraphs of the user’s own writing. It checks for 33 patterns, including significance inflation, promotional language, vague attributions, formulaic phrasing, AI-associated vocabulary, passive or subjectless fragments, dash overuse, chatbot closers, and excessive hedging. It produces a rewrite, audits the result for remaining AI signals, and applies a second rewrite.
- A content writer wants to reduce obvious AI-writing signals before publishing a draft.
- An editor needs to turn promotional, formulaic, or vague prose into more specific writing.
- A user wants an AI draft to better match their own sentence rhythm, word choices, and quirks.
- A user of Claude Code or another skill-compatible agent harness wants a plain-Markdown text-editing skill.
Pros & cons
- Documents 33 AI-writing patterns with before-and-after examples.
- Supports voice calibration from user-provided writing samples.
- Includes a final audit and second-pass rewrite.
- Uses plain Markdown and is described as portable across skill-compatible harnesses.
- Released under the MIT license.
- SKILL.md was not cached in the supplied material, so its actual frontmatter, full instructions, and implementation details cannot be verified.
- The material does not document a test suite, quality metrics, or platform-specific validation.
- It focuses on rewriting text rather than fact-checking, adding sources, or performing specialist editorial review.
- Aggressively removing AI signals may change the original voice or wording, so users should review the output.
How to install
With the Skills CLI, run npx skills add blader/humanizer. To update an existing installation, run npx skills update humanizer. To install across supported harnesses, run npx skills add blader/humanizer --agent '*'. Claude Code users can use the documented plugin commands, or copy SKILL.md into the harness’s expected skill directory. The supplied material does not specify the exact skill path.
How to use
Invoke it through the harness’s skill interface, for example /humanizer, then paste the text. You can also make a direct request such as Please humanize this text: [your text]. For voice calibration, provide two or three paragraphs of your own writing before supplying the text to rewrite. The exact trigger depends on the harness.