Chen Dahuang open-sourced his anti-AI-tone writing setup as a Skill: pick one voice card with real human examples per genre, and keep only 20–30 banned words as a guardrail
A common anti-AI-tone fix is to paste a list of hundreds of banned words. Chen Dahuang (@realchendahuang) argues this backfires: with only prohibitions and no good writing to imitate, the model either swaps in a synonym or strips sentences down to a manual, and putting long-form and short-post rules in one prompt mixes their voices. His setup has three layers: one "voice card" per genre, with the positive style, genre-specific don'ts and two real human-written examples; read only the one card for this piece; and an outer list of just 20–30 of the worst words and patterns. Two hard rules sit on top: don't invent facts, and when editing a draft keep its structure. He stripped personal details and open-sourced it as open-writing-skill, with four cards (long articles, short posts and tool recommendations, community questions, lessons-learned notes) plus a single-file version for web chat; he suggests Gemini with thinking set to Low. We installed it into a Claude Code project folder with npx skills add and checked the entry file and the four cards; we did not write with it or compare before and after.
Visit site open-writing-skill ↗- Pricing
- Free
- Region
- Global
- Where it's from
- Chinese community
- Added
- Nov 12, 2026
- Tested on
- Nov 12, 2026
- Version we used
- open-writing-skill git 09015ae; installed with npx skills add (claude-code)
- Categories
- WritingProductivity
- Original author
- 陈大黄 @realchendahuang ↗
- Our test
- Partly tested
- Cost / time
- MIT, free, Markdown only. The model you write with (Gemini, Claude, etc.) is billed by your own plan or API.
The task
When drafting or editing with AI, remove stock phrases and report-speak and get text that sounds like how you normally write.
Steps
- Install it in a tool that supports SKILL.md (Claude Code, Cursor, Codex and others): npx skills add realchendahuang/open-writing-skill, or clone the repo into your skills folder.
- Pick two or three of your own best pieces, or originals by human writers you like. Use only human-written text; never use AI output as an example.
- Put long-form pieces in voices/deep-article.md and short posts in voices/short-post.md, and under each add two or three sentences on how it opens, how it gives specifics, and how it ends.
- When writing, name the genre so the model reads only that one card plus references/banned-words.md; when editing, ask it to keep the draft's structure rather than rewrite or compress it.
- If you only use a web chat, paste prompts/web-single-prompt.md into System Instructions and replace the two example passages with your own; the author suggests Gemini with thinking set to Low.
- Write a few pieces and add any recurring tic to banned-words.md, keeping the list to 20–30 entries.
What the author says
Chen Dahuang @realchendahuang (X article, 12 Nov 2026): this is his own daily writing setup with personal details removed, a generic skeleton with no code dependencies that you can switch to your own voice in 10 minutes. He cites a COLING 2025 study linking the surge in words like delve to RLHF rater preferences, and shows four before/after pairs from his own library; one under-50-character question post of three lines, he says, drew over 360,000 views and 280+ replies. He recommends Gemini models with thinking set to Low for everyday writing, arguing that high reasoning tends to produce formulaic prose.
Our test
On 12 Nov 2026 on Linux we git-cloned https://github.com/realchendahuang/open-writing-skill (09015ae, MIT). In a temp folder, npx -y skills add realchendahuang/open-writing-skill -a claude-code -y installed it to ./.claude/skills/open-writing-skill. SKILL.md is 31 lines with name open-writing-skill in the frontmatter; it tells the model to read only one card per genre, check banned-words.md before delivering, and return the draft in a text code block. The four cards under voices/ (deep-article, short-post, probe-question, craft-notes) match the README; references/banned-words.md has three parts (stock words, formulaic patterns, ending rules); prompts/web-single-prompt.md is the web single-file version. We did not draft or edit anything with a model and did not evaluate the effect.
What we found when we tried it:The install command works and the repo matches the article: a short entry file, with the substance in four replaceable voice cards, so making it yours means editing those Markdown files.
How to get in
The original is a Chinese X article: https://x.com/realchendahuang/status/2120555090545459548 (use browser translate). The voice cards and examples are in Chinese; the banned-word list includes an English section (delve, tapestry, testament…), and English writers would need to replace the example passages with English ones.
Who it suits
People who often draft articles, posts, retrospectives or tool recommendations with AI and have a few past pieces they're happy with. Without your own examples, it does much less.
Pitfalls
- The cards ship with the author's own examples; unless you swap in yours, the output sounds like him.
- The author warns that using AI-generated text as examples makes the output more machine-like.
- Loading all four cards at once mixes voices; tell the model which genre this piece is.
- We did not write with it or compare it with similar skills such as humanizer; the effect is as the author describes.
More in this category
- Read the code, then fix calqued jargonCodingclear-namingCodexClaude CodeSkill free; no extra runtime◐ Partly testedChina
- 11 skills grow personal files into a queryable knowledge baseProductivityKnowledge baseClaude CodeWorkBuddyKit free (BY-NC); agent/Feishu extra◐ Partly testedChina
- Markdown to an X Article draft in one clickWritingMarkdown to X ArticleChromeClaude CodeExtension free; Articles need X Premium◐ Partly testedChina
- One sentence turns agent copy into printable PDFWritingKamiClaude CodePDFSkill free; needs a coding-agent plan◐ Partly testedChina
- One topic in, a sourced explainer short outVideoClaude CodeXiaomi MiMoRenoiseBase tier free; needs Claude plan◐ Partly testedChina