Humanizer Skill: Edit AI Prose while Preserving Facts
A passage can be grammatically correct yet sound unlike anything its author would say. Humanizer structures prose editing around the writer’s voice without adding facts to make the piece sound more convincing.

Humanizer edits style while preserving substance
Blader's Humanizer is an Agent Skill for checking and correcting text that has repeating patterns in AI writing. Examples include broad openings before the point, emphasizing importance without information, repeating sentence patterns, or turning ordinary language into advertising. The skill supplies a workflow for an existing AI assistant, not a new language model or authorship detector.
The source for this article is blader/humanizer identifies version 3.1.0 at commit 8c307f36 Several developers use the name humanizer; check skill and repository ownership before installing. This version prioritizes author voice and facts over stylistic deletion rules. Use writing samples to guide revisions when available.
It suits articles, emails, reports, and explanations whose substance is complete but whose language does not fit the reader. Correct wrong figures or missing evidence first. Polishing an incorrect statement can make it more readable without fixing the error.
Read the whole piece, rewrite, then compare the changes
The original workflow reads for stiff or formulaic wording, rewrites while preserving facts, compares against the original, and produces a natural final version. Read whole paragraphs: structural problems such as repeated closing sentences can exist even if individual sentences seem fine.
Words on Humanizer’s watch list do not always need to be removed. Connectors, repetition, and comparisons may convey meaning or reflect the writer’s voice. Check the context before editing and retain the register the work needs; a formal document may still need formal language.
Install only selected skills and supply enough context
The Skills.sh listing displays the command below for this source. Node.js and a supported agent environment are prerequisites. Review repository files and the scope of files to be added before confirming installation. This article did not run the installation command.
npx skills add https://github.com/blader/humanizer --skill humanizer
After installation, confirm the agent can see the skill and invoke its supported name, for example, “Use humanizer to edit the following paragraph.” Slash-menu invocation varies by app. Provide the original, audience, purpose, formality, and terms that must remain unchanged.
Example hypothetical request: Use humanizer to edit this announcement for a Thai customer. Use friendly and to-the-point language. Keep the service name, date, terms, and links intact. Do not add testimonials or personal experiences. Send a revised version and indicate where the meaning might change.
Thai example: remove filler while keeping the original conditions
The following hypothetical example explains the principle. It has not been tested with Humanizer and does not guarantee the actual service's capabilities.
| Edition | Text |
|---|---|
| Before editing | When every second matters, our tools do more than combine files: they help improve efficiency by letting you merge PDFs in your browser |
| After editing | Merge PDF files in the browser with this tool. |
The removed wording merely inflated importance without adding usage instructions. The remaining claim is that PDFs can be merged in the browser. If the original does not say offline, free, or unlimited file size, an editor has no grounds to add those terms even if they make the sentence more appealing.
Also watch confidence wording. Changing “may reduce time” to “reduces time” strengthens the claim. Shortening “within 7 business days after all documents are received” to “within 7 days” removes conditions. When summarizing, preserve both conditions and confidence levels.
Check facts that must be preserved separately from tone
KrobClick suggests two review passes. First check substance: names, organizations, figures, units, dates, conditions, negations, links, quotations, and certainty. Then read aloud for rhythm, repetition, and audience fit, preventing fluent prose from concealing lost information.
- If the author expresses uncertainty, does the new version preserve it?
- Are necessary technical terms explained with evidence, or replaced in ways that change meaning?
- Has the revision added personal experience, opinions, or testimonials the author did not supply?
- Were code-block text, commands, paths, or metadata changed unintentionally?
The source provides a file mode for editing prose while retaining code, commands, paths, metadata, and link destinations. Review the diff before saving or publishing a whole-file edit, especially when commands appear beside explanatory text.
Do not use AI-detector scores as the measure of success
The referenced skill does not prove outputs pass any AI detector, and this article did not test detectors. Evaluate factual accuracy, preserved author voice, and reader understanding, and follow organizational disclosure requirements rather than using edits to evade them.
Thai rhythm and sentence endings differ from the skill’s English examples. Removing dashes or changing connectors alone will not solve every problem. Start with a short paragraph and a clear audience, then compare the revision’s facts and voice with the original. Extend the process to the whole document and use the Inspect work from AI Agent guide to check the actual work before publication.
Sources
Source reviewed October 21, 2026, pinned to cited commits. This explains documentation without installing or testing the skill. Thai examples are hypothetical KrobClick scenarios. The AI-generated cover is not output from the described skill.