Huorengan

Audit and rewrite Chinese or English content to remove AI tone ("AI-isms"), then pull it toward a target human voice. Use this skill when asked to "remove AI tone," "sound human," "去 AI 味," "说人话," "活人感," "make this sound like a real person," or "less like a template." Bilingual (zh/en), with detect-only, edit-in-place modes, scene packs, protected spans, and voice profiles.

活人感 huorengan — Audit, Rewrite & Re-voice / 审稿、改写与注入人声

Subtraction + addition. First remove the "dead" (AI tone), then inject the "alive" (a target human voice). 减法 + 加法：先去掉「死」，再注入「活」。

⚠️ Stage 0 skeleton. This file is a placeholder so the CI loop (count consistency, plugin sync, policy alignment) closes end-to-end before the full bilingual rule set lands in later stages. Each ### below is a real detection category slot that the engine will implement.

What this skill is and isn't / 这个 skill 是什么、不是什么

This is a writing-quality tool, not a verdict. The patterns flagged here are statistically more common in LLM output, but humans on autopilot — writing under deadline pressure, in unfamiliar genres, or in a second language — produce the same shapes. Signals, not proof.

这是一个写作质量工具，不是判决。这里标记的模式在 LLM 输出中更常见，但赶稿的人、写不熟悉的体裁、或用第二语言写作的人，也会写出同样的形状。是信号，不是证据。

Modes / 模式

• rewrite (default / 默认) — Flag AI tone and rewrite, then (if a voice is set) pull toward it. 标记并改写，再按目标人声拉拢。
• detect — Flag only, no rewriting. 只标记不改写。
• edit — Edit a file in place with minimal, targeted edits. 就地最小改动。

Every mode runs protected-spans detection first — version numbers, commands, paths, errors, quotes must not drift. 所有模式第一步先做 protected-spans 识别。

Working style / 交互方式

Work like an editor, not a slogan filter. The goal is not just to delete AI-looking phrases, but to help the user end up with text they can actually send.

像编辑一样工作，不像关键词过滤器。目标不只是删掉 AI 套话，而是帮用户拿到一版真的能发出去的文本。

Default flow:

1. Fence protected spans first — numbers, dates, commands, paths, quotes, owners.
2. Name the main problem — vocabulary, structure, translation tone, rhythm, or voice drift.
3. Choose the lightest effective move — cut the pose, keep the information.
4. Do one quick second pass — check for opener residue, summary residue, narrator residue, empty judgment, and over-even rhythm.
5. Return something usable — either a rewrite, an audit, or minimal in-place edits, depending on mode and scene.

Default principles:

• Fidelity before smoothness.
• Name the problem before changing the sentence.
• Use the smallest stable edit that solves it.
• Do not invent facts, sourcing, or attitude just to sound more human.
• Be direct when the scene wants directness; be conservative when the scene carries risk.

What to remove or fix / 该删除或修改的

Detection categories below. Each ### is one category that the engine implements and CATEGORIES.md maps to a detector type. Bilingual where the rule applies to both languages; language-specific where noted. Full phrase lists live in references/; the engine implements the regex-detectable subset (see detector/CATEGORIES.md).

Tier 1 vocabulary (always flag) / 一级词汇（默认替换）

Words 5–20x more frequent in AI text. Replace on sight. zh: 开场套话（值得注意的是/综上所述）、商业黑话（赋能/抓手/闭环）、小红书腔（保姆级/绝绝子）、调试腔（兜住/收口/根因）、谄媚（好问题/稳稳接住）、价值拔高（不仅仅是…更是）、无源引用（研究表明）、正能量收尾（未来可期）。en: delve, tapestry, leverage, seamless, robust, comprehensive, game-changer, "serves as", "at its core". See references/phrases-zh.md, references/phrases-en.md.

Tier 2 vocabulary (flag in clusters) / 二级词汇（同段聚集才标记）

Individually fine; 2+ (short para) or 3+ (long para) in the same paragraph is the signal. zh: 然而/此外/与此同时/显著/有效/全面/持续 + 单音节命令词（补/接/核/进/顺/落/坏/跑）。en: harness, navigate, foster, elevate, nuanced, crucial, transformative, cornerstone. Keep the best-fit one, rewrite the rest.

Tier 3 vocabulary (flag by density) / 三级词汇（全文密度过高才标记）

Common words; only flag when saturated (~3%+ of text). zh: 重要/关键/核心/创新/优化/提升/推动/确保。en: significant, innovative, effective, dynamic, compelling, unprecedented. Replace some with specifics (numbers, names, examples), not all.

Structural anti-patterns (cross-lingual) / 结构反模式（跨语言）

19 shapes from references/structures.md: binary contrast (不是X而是Y / "It's not X, it's Y"), summary closer (综上所述 / "In conclusion"), mechanical ordering (首先…其次…最后), symmetry padding (既要…又要), value inflation, positive-energy closer, psych judgment. Cross-lingual types share one type so bilingual symmetry holds.

Translation tone (Chinese-specific) / 翻译腔（中文特有）

zh-only types — English-thinking literally translated into Chinese. 被动语态堆砌（被…被…被…）, 长定语结构（的…的…的…）, 「基于…」开头, 「通过…来…」. No English counterpart. See references/translation-tone.md.

Chatbot artifacts / 机器人痕迹

"I hope this helps!", "Great question!", "Certainly!", "Let's dive in!" / 好问题！希望这对你有帮助！让我来为你解释. Remove entirely. Also reasoning-chain artifacts ("Let me think step by step", "Here's my thought process").

Significance inflation / 意义拔高

"marking a pivotal moment", "a watershed for the industry" / 深刻的影响, 前所未有, 颠覆性变革, 范式转移. State what happened; let the reader judge significance.

Vague attribution / 无源引用

"Experts believe", "Studies show", "Research suggests" / 研究表明, 数据显示, 业内人士认为, 有专家指出. Either cite a specific source or drop the attribution and state the claim directly. Don't fabricate sources.

False-concession structure / 虚假让步

"While X is impressive, Y remains a challenge" / 虽然…但是…. Either make both halves specific or pick a side. The balanced-sounding non-statement is the tell.

Promotional language / 推销腔

"nestled within breathtaking foothills", "a vibrant hub of innovation" / 打造, 助力, 全方位, 深度赋能. Replace with plain description. If you wouldn't say it in conversation, cut it.

Social endorsement closers / 社交式收尾

"This one is worth your time:", "do yourself a favor and read this", "thank me later" / 建议收藏, 强烈推荐, 划重点. Say what the thing is and who it's for; drop the generic endorsement.

Hedge-stacked predictions / 对冲堆叠

"could potentially create", "may eventually unlock" — modal + hedge adverb stacks. Each hedge cancels the next, asserting nothing. Pick one.

Formulaic openers / 公式化开场

"In the rapidly evolving world of X…" / 在当今…的时代, 随着…的不断发展. Lead with the news/insight; context can come second.

Emotional flatline / 情感平淡

"What surprised me most", "the most interesting part" / 你不是敏感, 你只是太久没被稳稳接住了. Tell-don't-show. If the emotion is real, the writing earns it; if not, cut the claim.

Novelty inflation / 新颖性拔高

"the failure mode nobody's naming", "a concept nobody talks about" / 真正的X不是…而是…. Assume the concept isn't novel and frame accordingly.

AI-tool fingerprints (placeholders / citations / UTM) / AI 工具指纹

Near-definitive single-hit signals: unfilled [Your Name] placeholders, citeturn0search0 citation markup, utm_source=chatgpt.com. Strip mechanically. Each is proof the text was copy-pasted from a specific chat tool.

Rhythm & uniformity (stylometric) / 节奏与均匀度

Structure is the #1 detection signal. Sentence-length uniformity (CV < 0.25), uniform paragraph length, low TTR (< 0.40 at 200+ words), punctuation-density uniformity across paragraphs, cross-paragraph burstiness. zh: 句长集中在 N 字（变化小）. Fix by mixing short punchy sentences with longer ones — vary, don't sand.

---

Protected spans (protect first) / 保护片段（先保护）

Before any rewrite — every mode — fence off what must never drift. See references/protected-spans.md.

• Numbers, dates, ranges, units / 数字、日期、版本号、区间、单位 — 不改数值，不四舍五入
• Names + attribution / 人名、组织、产品、issue/PR 编号、责任主体 — 不换主体，不模糊"谁做的"
• Quoted text + titles / 引号内原文、文章标题 — 默认原样保留
• Commands, code, params, fields, paths / 命令、代码、接口名、字段、路径 — 拼写大小写符号全保留
• Errors, logs, statuses, metrics / 报错、日志、HTTP 状态码、指标 — 不换错误类型，不丢范围

When a sentence can only sound natural by changing a protected span, keep the span, accept the stiffness. Fidelity wins over style.

---

★ Voice (addition layer) / 加法层（注入人声）

When voiceMode ≠ none, after subtracting AI tone, pull the result toward a target human voice. This is huorengan's step beyond both parent projects. See references/voice-contract.md.

Profiles (from policy/voice.toml): casual / professional / technical / warm / blunt / custom (calibrated from a sample).

The engine computes voice.drift (0-100, distance from target) and concrete suggestions: "split sentence 3 at word 15", "mix in 3-8 word punchy sentences", "swap to target connectors". zh: 「第 N 句约 X 字，考虑在 Y 字处断开」.

Hard boundary: voice suggestions must not touch protected spans. When voice pulls conflict with fidelity, fidelity wins. voice.drift is independent of score — a text can be clean (low score) yet far from a target voice (high drift).

Mode behavior / 模式行为

rewrite

• Default for pasted text, drafts, blurbs, release copy, README intros, and issue replies.
• Output a usable rewrite, not just a diagnosis.
• Keep the original scene and factual boundary intact.

detect

• Use when the user wants a read, review, or confidence check before rewriting.
• Group issues by severity and distinguish clear problem from judgment call.
• If sourcing or fidelity is the main risk, say so plainly instead of force-rewriting.

edit

• Use the smallest change set that solves the problem.
• Do not reorder paragraphs, merge nearby sentences, or rewrite whole sections unless the user asked for that level of change.
• For code-adjacent text, only touch the comment/docstring/message text, never the code-bearing span.

Second-pass audit / 二次回读

After the first rewrite, always do one quick residual check. Look for only five things:

1. opener residue (值得注意的是, 直接说结论, Great question)
2. summary residue (综上所述, 总的来说, In conclusion)
3. narrator residue (更重要的是, 这说明了, what this shows is)
4. empty judgment (意义重大, 方向是对的, pivotal, transformative)
5. over-even rhythm (too many sentences with near-identical length)

If the first pass already protects facts and reads naturally, keep the second pass light. Do not polish the life out of it.

---

Output format / 输出格式

rewrite mode

1. Main issues found (quote the concrete trigger where useful)
2. Rewritten version
3. What changed and why
4. Second-pass audit
5. Voice drift notes (if voice set)

detect mode

1. Issues found (grouped P0/P1/P2)
2. Assessment (clear problem vs. judgment call)
3. If relevant: whether the safer move is audit-only instead of rewrite

edit mode

1. Edits made (file location + before → after)
2. Verification
3. Whether any protected spans were intentionally left stiff for fidelity