The Origin: I Spent Over Ten Hours Testing Ten "De-AI Flavor" Skills
First, here is the open-source address for [Human Talk.skill]: https://github.com/Pluviobyte/rnskill
Recently, various "De-AI flavor" Skills and open-source projects have become very popular in the Chinese community. A quick search on GitHub for humanizer, shuorenhua, stop-slop, qu-ai-wei, De-AI-writing... and you can find over a dozen.
My needs are very specific: when writing AI-related technical articles, I want the draft to read more like I wrote it myself, with less of a template feel and "AI flavor." So, I cloned all the projects I could find to my local machine and ran a round of comparative tests using the same short draft on an AI Native theme.
After testing ten projects, my first discovery was: they are not the same type of De-AI Skill at all. I was misled by clickbait bloggers again.
Three That Can Be Used Directly for Chinese Technical Drafts
shuorenhua, Humanizer-zh, and De-AI-writing.
shuorenhua has the finest perception of scene and register. It first determines whether your text is a technical review, an opinion piece, or documentation before deciding what to change and what to keep. After the revision, terminology and judgments usually remain, while the template shell and empty summaries are cleared away.
Humanizer-zh has the broadest rule coverage, with corresponding treatments for 24 types of AI writing traces. It's suitable for a first round of cleaning, but it occasionally turns the draft into something resembling a generic press release polished by an editor, weakening the personal voice.
De-AI-writing makes the lightest changes. It prioritizes preserving the original structure, only clearing out signpost words, lecture-style tones, and paragraph-ending summaries. It's suitable when you are relatively satisfied with the original draft and worry that major changes might go off track.
Two Suitable for Extracting Rules
stop-slop and the original English version of humanizer.
stop-slop's rules are short, hard, and direct. For example, it targets binary contrast sentences ("XX is not A, but B"), triple parallelism, aphorism-style paragraph endings, and over-explanation for deletion. Applying it directly to Chinese is a bit stiff, but it's excellent when broken down into a "forbidden list."
The original English humanizer is the most systematic, categorizing AI traces into four major types: content patterns, grammar, style, and communication patterns. Chinese adaptation needs to be done manually, but it is highly valuable as an upstream reference.
Two for Long-Term Writing Stability
writing-agent is a complete writing pipeline. It covers the entire process from topic selection, evidence collection, and stance confirmation to proofreading, de-flavoring, and exporting. I ran workflow validations and 15 unit tests locally, and they all passed. The full process requires configuring models and APIs, making it suitable for future research into long-term official account writing.
nuwa-skill focuses on style distillation. It requires feeding 5-7 of your real articles to let it extract your writing characteristics and generate a personalized Skill. Testing it with a single paragraph isn't very meaningful.
Three That Are Far from "De-AI Flavor for Chinese Technical Articles"
chatgpt-comparison-detection is a repository for the HC3 dataset and detection research. I ran sample text using its built-in Chinese instruction list, and it hit one high-frequency ChatGPT word: "so." It's for detection research, not a revision tool.
ai-flavor-remover is a standalone prompt without a Skill structure, suitable for throwing directly into reasoning models to try out.
taste-skill is a front-end aesthetic Skill that manages interface design and has nothing to do with text.
What I Discovered Later
The most useful output isn't the final draft from a specific Skill, but the specific rules extracted from different Skills.
shuorenhua made me notice the issue of scene switching. Writing a Fable-5 review and writing an AI Native opinion piece requires keeping different things. In a review, hard data like price, speed, and model comparisons cannot be touched; in an opinion piece, personal judgments and experiential details shouldn't be smoothed over.
stop-slop made me realize how frequently the "XX is not A, but B" sentence structure appeared in my own drafts. Once noticed, it becomes hard to tolerate. Similar issues include empty summary endings, three-part parallelism, and transition filler like "in other words."
Humanizer-zh's binary contrast detection and De-AI-writing's fidelity strategy can also be extracted for individual use. The former helps me identify sentence structure issues, while the latter prevents over-editing from scattering terminology and judgments.
How I Use It in the End
I didn't choose one "strongest" Skill to install; instead, I compiled my own checklist from these projects. After writing a technical article, I go through it:
Does the draft contain actual testing experiences and my own judgments? Without these, the article becomes a generic manual that anyone could write, and readers won't know it came from a specific test.
Are there binary contrast shells, empty summaries, triple parallelism, or aphorism endings? Delete them on sight.
Have terminology and model names been scattered? If Fable-5 is changed to "this model" or Claude Opus is changed to "this product," that's over-editing.
Are the sentence lengths too uniform? If every sentence is roughly the same length, it reads very flat, as if it were uniformly polished.
These rules are more effective than any single Skill. Everyone's writing style is different, and the source of "AI flavor" varies. Breaking down the rules and combining them yourself works better than directly applying a generic Skill.
Finally, the Best Matching Model—Opus 4.6
During my testing, the GPT series performed the worst, especially when used in Codex, as it often missed Skill rules. DeepSeek V4 Pro performed better. Opus 4.8 likely distilled GPT's style and also performed poorly. The best-performing model was Opus 4.6.
The above article was produced using Human Talk Skill and is open-sourced: https://github.com/Pluviobyte/rnskill
Welcome to follow me @Pluvio9yte. In the next issue, I will explain the detailed usage of this Skill.
Next issue preview: "1,000 Followers in 7 Days: I Broke Down Video Production into an AI Production Line"
