De AI Writer

De-AI Writer — Chinese AI-Smell Removal

Overview

Chinese AI writing has its own tells, and they are not the English ones. English humanizers hunt delve, "it's not just X, it's Y" and em-dash overuse; a Chinese draft reads machine-written because of 赋能 / 闭环 / 抓手 / 底层逻辑 (pattern 15), 首先-其次-最后 scaffolding (pattern 30), 随着…的发展 openers (pattern 25), 拔高意义 endings (pattern 16), and 公文套话 (pattern 22). A translated English humanizer misses all of it.

This skill ships the pattern catalog plus the editing procedure: score a draft for AI smell, rewrite it against the specific patterns it hits, clone a reference style, and review the result. The full 35-pattern catalog lives in references/ai-patterns-zh.md and is plain Markdown — usable as a prompt by any assistant.

Note on the engine: this bundle is documentation only. A zero-dependency local rule engine is published in the source repository (see source_repo) as an optional external prerequisite — it is not included here.

When to Use This Skill

• Use when a Chinese draft "reads like AI" and needs to sound human-authored.
• Use when the user asks 去AI味, 改得像人写的, or 这段是不是AI写的 (is this AI-written?).
• Use when editing marketing copy, WeChat articles, product listings, or social posts written in Chinese.
• Use when asked to imitate a reference writing style (风格克隆) or to produce A/B variants of the same copy.
• Use when shifting tone: casual / formal / marketing / humor / direct.

How It Works

Step 1: Score the draft first (diagnose before editing)

Scan the text, record which patterns hit and where, then compute the AI-smell index with the deterministic formula below. Do not rewrite from vibes — the same phrases recur, and you need the hit list to verify the edit afterwards.

The AI-smell index (deterministic — the same formula must be used before and after)

1. Count hits per paragraph. For each pattern, count one hit per paragraph — repeats inside the same paragraph do not inflate the score.
2. Weight by evidence strength. 强模式 = 2 points; patterns the catalog marks 弱证据 (破折号 / 限定词 / 被动与无主语 / "的"-字堆叠 / 引号不统一) = 1 point.
3. Normalize by length. D = 加权总分 / max(1, 总字数 / 100) — weighted hits per 100 characters.
4. Index. AI味指数 = min(100, round(D × 10)).
5. Bands. 0–20 基本像人写 · 21–45 轻度 AI 味 · 46–75 明显 AI 味 · 76–100 一眼假.
6. Report three numbers, not one: 命中处数 / 加权总分 / AI味指数. After rewriting, recompute with the same formula so the delta is comparable. If a hit cannot be attributed to a catalogued pattern, report only the observable hit count and say the index is not computed — never invent a number.

Worked example (the sample below): 8 hits in 77 characters, 6 strong (6 × 2 = 12) + 2 weak (2 × 1 = 2) → 加权总分 14 → D = 14 / 0.77 ≈ 18.2 → 指数 = min(100, 182) = 100/100.

AI 味体检报告
总字数 77 ｜ 命中 8 处 ｜ 加权 14 ｜ AI味指数 100/100（一眼假）
机械连接 ×3   官方黑话 ×2   空洞拔高 ×2   夸张词 ×1

Step 2: Rewrite against the specific patterns, not in general

Work through the hit list one pattern at a time. Correct each pattern by its own rule (the catalog gives 识别特征 → 为什么假 → 改前/改后 for all 35). Two rules govern the whole pass:

• A single hit is not evidence. The catalog marks certain patterns (em-dash, hedges, passive voice, 的-stacking, quotation marks) as 弱证据 — only act when two or more appear in the same paragraph.
• Delete, don't decorate. Most patterns disappear by deleting the sentence that carries them: drop the negation half of 不是 X，而是 Y, drop the significance ending, drop the 开场铺垫, drop the assistant residue (希望对你有帮助).

The facts come from the source, never from the pattern list. A rewrite may delete packaging, reorder, and rephrase; it must not introduce a fact the source does not contain. If the rewrite would be clearer with a number or a specification (rotational speed, battery life, materials), and the source has none, ask the user for it — never supply one.

Step 3: Verify the rewrite

Re-score the rewritten text with the same Step 1 formula. The index should drop and the semantic content must be preserved — report the before/after triple (命中处数 / 加权总分 / 指数), how many patterns were cleared, and confirm that no fact was added or lost. A rewrite that lowers the score by deleting facts is a failed edit; so is one that raises the count of facts in the text.

Step 4: Optional — style clone, variants, tone, review

• Style clone: supply a reference sample (an old article, a novel fragment, a writer you like) and match its sentence rhythm, vocabulary and colloquial ratio.
• Variants: produce 2-6 clearly different versions (short and punchy / loose and spoken / vivid) for headline and ad-copy A/B tests.
• Tone: re-target the same content to casual, formal, marketing, humor or direct register.
• Review: score the finished text on Hook / Pacing / Emotion / AI-Smell / Clarity / Persuasion / Structure / Readability, plus three concrete improvements.

Examples

Example 1: Business copy with stacked AI tells (deletion-only edit)

Before:

首先，这款产品不仅性能卓越，更是彰显了品牌的匠心精神。随着智能家居时代的发展，
它以用户为中心，赋能每一个家庭的智慧生活，具有里程碑意义。

After — deletion-only; every surviving phrase already appears in the source, and no specification was added:

这款产品性能卓越，能融入家庭的智能生活。

Every change is tied to a catalogued pattern:

• 首先 → pattern 30 (强制编号与小标题, mandatory 首先/其次 numbering)
• 不仅…更是 → pattern 1 (不是 X，而是 Y — negation-half padding)
• 随着智能家居时代的发展 → pattern 25 ("随着……的发展"式开头)
• 以用户为中心 / 赋能 → pattern 15 (高频 AI 词) and pattern 23 (互联网黑话)
• 彰显了品牌的匠心精神 / 具有里程碑意义 → pattern 16 (拔高意义)

Six patterns removed, zero facts added, zero facts lost: the rewrite keeps 性能卓越 and 智能生活, the only two claims present in the source. It does not gain a rotation speed, a battery life or a waterproof rating — those are not in the source, and inventing them is exactly the failure mode this skill exists to prevent.

Example 2: Assistant residue in generated output

Before:

当然可以！以下是我为你整理的三个要点，希望对你有所帮助。需要我展开说明吗？

After:

三个要点：

Pattern 31 (聊天机器人残留) — the wrapper adds nothing; the content stands alone.

Best Practices

• ✅ Score before and after; report the triple (命中处数 / 加权总分 / 指数) so the edit is verifiable.
• ✅ Quote the pattern number for every change — it makes the edit reviewable and teachable.
• ✅ Keep every fact, number and claim from the source; only the packaging should change.
• ✅ Respect 弱证据 — two hits in one paragraph, not one hit anywhere.
• ✅ Keep the catalog in Chinese when editing Chinese; the tells are language-specific.
• ❌ Don't invent facts to make the rewrite concrete — if the source has no number, ask for one.
• ❌ Don't swap one AI word for another AI word (赋能 → 助力 solves nothing).
• ❌ Don't "polish" a draft into formal register — that usually adds AI smell rather than removing it.
• ❌ Don't strip caveats and qualifiers that carry real meaning (legal disclaimers, safety warnings).

Limitations

• Chinese-centric by design. The pattern catalog targets Chinese AI tells; it will not fix English AI smell (use an English humanizer for that).
• Documentation only. No script and no engine is bundled in this skill. The optional local rule engine is published in the source repository; deep rewriting, style cloning, variants and scoring require an LLM of your choice.
• Rule-based detection, not a detector model. The index is a heuristic over known patterns, computed by the formula above. It cannot prove authorship and must not be used as evidence that a text "was" or "was not" AI-written.
• Unlisted patterns are out of scope. Tells that are not in the 35-pattern catalog pass through untouched; the catalog is the ceiling of what this skill sees.
• Keep-conditions need human confirmation. Some removal is context-dependent — a safety caveat such as "do not soak for long periods" may be a legal requirement; when a pattern overlaps with a claim that must stay, ask before deleting.

Reference

• references/ai-patterns-zh.md — the full 35-pattern Chinese AI-smell catalog, grouped into 摆姿势 / 机械节奏 / 注水借势 / 格式装饰 / 助手残留, each entry giving 识别特征 → 为什么假 → 改前/改后.
• Source repository (MIT, optional runnable engine): https://github.com/jiawood2006/hermes-skills