anti-vibe-writing

anti-vibe-writing

本 README 提供中英文两个版本：English（本页，默认）和 中文。点上方徽章切换。

One goal: make AI sound genuinely idiomatic, 倍儿地道.

An agent writing skill that removes the AI-generated feel from documents. Works in English and Chinese, with optional modes for matching a specific author's voice.

It runs best as a final pass after drafting with Claude Code, Codex, or any LLM-backed agent. The goal is not to make the prose prettier. The goal is to keep the substance and remove the tells: templated phrasing, vague abstraction, consultant-speak, markdown-heavy formatting, over-structured outlines, the copy-paste residue chat and search models leave behind (oaicite / turn0search0 / [cite: 1] / stray 【】 markers), and the cautious over-balancing that makes writing feel assembled.

If the default "clean" mode is not enough, the skill also supports:

• Human-texture mode: inject controlled irregularities (inversions, particles, half-sentences, non-standard punctuation) for personal voice
• Learning mode: build a reusable host profile from real samples so future drafts sound like the person who would write them
• Scenario presets: format and tone constraints tuned for tweets, Weibo, blogs, podcast show notes, and professional reports

Quick start

At its core this is just a set of markdown rules (SKILL.md + references/ + assets/), not tied to any one agent. Claude Code, Codex, Kimi, work-buddy, Hermes. If it can read files or take an instruction prompt, it can use this.

1. Get the rules

Clone the repo, or just copy the skills/anti-vibe-writing/ folder:

git clone https://github.com/weijt606/anti-vibe-writing.git

2. Feed it to your agent (pick whichever fits)

• Simplest: just hand it the repo link. No need to clone first: send https://github.com/weijt606/anti-vibe-writing and let the agent read SKILL.md and references/ and configure itself. Works with any agent that can browse the web or run git.
• One-off (any agent): open skills/anti-vibe-writing/assets/rewrite-prompt-template.md. It has ready-made instruction blocks: Full Rewrite / Light Cleanup in English, and "中文改写（带负向约束）" for Chinese. Copy the block for your language and send it with your draft.
• Persistent: put SKILL.md and the matching references/*patterns-to-remove.md wherever your agent loads context. Names differ by tool:
  • Agents with a skills directory (e.g. Claude Code): drop it in ~/.claude/skills/anti-vibe-writing/, then call /anti-vibe-writing
  • Agents with a project-instructions file (e.g. Codex's AGENTS.md): write or include the rules there
  • Otherwise: paste into the system prompt / custom instructions / knowledge base

3. (Optional) Name the scenario and mode

One line of context changes the result a lot:

• Scenario: "this is a tweet / a newsletter / a technical memo"
• Loosen up: "make it feel personal" / "blog voice" → enables human-texture mode
• Match a voice: paste a few of your own samples and say "learn my style" → enables learning mode

When in doubt, say nothing. The default clean mode is right for most drafts.

Quick examples

Curated before/after snippets live in examples/ for anyone scanning the repo. The full regression set stays under references/.

Voice modes

The skill runs in one of three modes:

Modes combine with scenario presets (tweet / Weibo / blog / podcast / report). Conflict resolution rules are documented in SKILL.md.

Human-texture mode also has a social-only "casual typing" (随手打) layer (off by default; turn it on by naming it, "casual typing" / "开随手打", or describing the effect, "like a quick phone post"; matched by intent, not a fixed phrase, and a plain "loosen it up" won't trigger it): a tiny amount of phone-typing texture on casual posts (dropped end punctuation, no capitalization, an omitted particle) that never touches numbers, names, or links and never makes meaning-changing typos. It's phone-typing texture, not error injection to dodge AI detectors.

Use cases

The flagship case: Chinese posts on X. Chinese AI-smell is most obvious on X: 赋能 / 打通, 首先 / 其次, three-clause parallelism, and machine-translation syntax give it away at a glance. This skill is built for exactly that: take an "obviously AI-written" Chinese post and make it read like something a person actually typed. See examples/07-tweet-zh.md and examples/08-translationese-zh.md.

Other common cases:

• Social posts (X, Weibo, Jike, RedNote)
• Blogs, newsletters, public WeChat articles
• Podcast show notes and video scripts
• README cleanup
• Product docs and landing page copy
• Proposals, founder notes, technical memos, internal reports

Output goals

• More human rhythm
• More intentional structure
• Cleaner phrasing
• Stronger voice
• Less AI smell
• Sounds chosen by a person, not assembled by a system

Repository layout

agents/
  README.md
  anti-vibe-writing-dev.agent.md           # local, gitignored
  anti-vibe-writing-dev.agent.example.md
skills/
  anti-vibe-writing/
    SKILL.md
    references/
      patterns-to-remove.md                # English AI-smell
      chinese-patterns-to-remove.md        # 中文 AI 味
      before-after-benchmarks.md           # English benchmarks
      chinese-before-after.md              # 中文基准
      common-problems-and-fixes.md
      human-passes.md
      human-texture.md                     # Optional irregularity
      learning-mode.md                     # Sample-driven style learning
      scenario-presets.md                  # Per-scenario constraints
    assets/
      final-pass-checklist.md
      rewrite-prompt-template.md
      host-profile-template.md             # Fillable host profile
      style-extraction-prompt.md           # One-shot extraction prompt
examples/
  ...
CHANGELOG.md
CONTRIBUTING.md
README.md                                    # English (default)
README.zh.md                                 # Chinese
LICENSE

Working with the files

Skill files are versioned. The developer agent file is local and gitignored by default, so contributors can adjust it without changing the public repository.

To create a local agent file, copy agents/anti-vibe-writing-dev.agent.example.md to agents/anti-vibe-writing-dev.agent.md.

If you want tool-specific auto-discovery, you may still need to mirror these files into the locations required by the target agent platform.

Credits & references

This skill stands on the shoulders of several open de-AI / humanizer projects and writeups. The patterns below were studied and adapted into this skill's own structure; the original analysis and wording belong to their authors. Thanks to:

English:

• blader/humanizer: a 30-pattern humanizer skill (MIT). Informed the sentence-level tells: copula avoidance, negative parallelism, synonym cycling, false ranges, signposting, diff-anchored writing.
• hardikpandya/stop-slop: AI-slop detection skill (MIT). Source of the optional five-dimension scoring pass (Directness / Rhythm / Trust / Authenticity / Density) in assets/final-pass-checklist.md.

中文 / Chinese:

• op7418/Humanizer-zh: a 24-pattern Chinese humanizer skill, itself a Chinese adaptation of blader/humanizer (MIT). Informed the copula-"是" avoidance and synonym-cycling tells in the Chinese track.
• "AI 中文翻译腔" by yage.ai: the analysis of Chinese translationese (物理动作动词写抽象 / 形容词加冒号预判读者 / 抽象名词主语). Informed the 翻译腔层 of references/chinese-patterns-to-remove.md.
• @dotey on X: discussion of de-AI prompt techniques (role-setting, negative constraints) that shaped the 改写心态 section and the Chinese rewrite prompt block.

These are independent projects with their own scope; this repo borrows ideas, not code. If you maintain one of them and want a credit adjusted, open an issue.

Contributing

See CONTRIBUTING.md for how to add new patterns, benchmarks, or scenario presets.

License

This project is open source under the MIT License. See LICENSE.

Contributor notes

• Preserve meaning. Sharpen the writing without changing facts.
• Prefer concrete edits over generic style advice.
• Keep structure only when it helps the reader.

Version highlights

1.7.0

• New copy-paste model-residue layer for the machine-only tokens current chat and search models leave in a pasted draft: contentReference / oaicite / turn0search0, [cite: 1] / [cite_start] / [span_1], grok_card, ppl-ai-file-upload, stray 【】/† citation scaffolding, unsourced [1][2] brackets, and leftover "Here is the revised version" / "Sources:" framing. Exact-string and language-agnostic, so it's added to both patterns references, the deterministic grep gate, the checklist, and the rewrite-prompt template
• Chinese track adds an 附和 / 谄媚开头 tell (问得好 / 你说得对 / 好的，下面是……) and an 无源的权威铺垫 tell (研究表明 / 数据显示 / 专家指出 with no source named)
• New English sentence-level tells (sycophantic openers, unsourced authority) and a note that the AI vocabulary list drifts by model generation (showcasing / highlighting / emphasizing / enhance now sit beside the older delve / tapestry set), so no single wordlist is treated as final
• Additive only: no change to the skeleton, philosophy, or workflow. Still a lightweight set of markdown rules

1.6.0

• The final checklist is now a step you run, not a list you glance at: after rewriting, work through final-pass-checklist.md, fix only the flagged spots, re-check, and stop after at most two rounds. It's the lightest form of a generate → check → revise loop. One model, one conversation, no extra agents, still just markdown
• New deterministic gate: a one-line grep at the bottom of the checklist catches the exact tells self-review skims past (em-dash —, the … character, stray → • in prose, and a fast subset of the jargon list). Shared double curly quotes “ ” are left to human judgment by language, so Chinese full-width quotes aren't deleted by mistake
• Learning mode gains a closing check: compare the output's sentence rhythm and punctuation against the numbers recorded in the host profile, and nudge it back if it drifted