remove-ai-flavor

name

remove-ai-flavor

description

Revise Chinese writing to remove AI-flavored, template-like expressions while preserving meaning, facts, and author voice. Use when the user asks to 去除AI味, 降低AI感, 改得更像人写, 润色小红书/公众号/随笔/文案/评论/邮件/小说正文, or specifically remove patterns like “不是...而是”, “先...再”, “真正...的是”, “这次只看”, “X很简单：”, mechanical contrast/sequence framing, over-neat parallel rhythm, repeated route markers, and unexplained rhetorical questions at the end.

Remove AI Flavor

Goal

Use this skill to make Chinese prose feel less machine-polished and more like a specific person wrote it. Keep the user's meaning, facts, stance, and original register. Remove only the structures that make readers feel the text is being assembled from an assistant template.

This is a preservation-first editing skill. Do not invent facts, examples, data, jokes, or personal experience just to make the text look human.

Default Workflow

1. Identify the text type:
   • Social posts: keep hook strength, density, and shareability.
   • Essays or comments: keep the author's stance and rhythm.
   • Academic or technical prose: keep terms precise and claims properly hedged.
   • Fiction: keep scene continuity, character action, dialogue intent, and manuscript-only output.
2. Scan the text in this order:
   • Structure: paragraph sameness, list-like expansion, repeated “claim + explanation + summary”, over-even sentence distribution.
   • Voice: assistant route markers, fake intimacy, excessive second person.
   • Sentence shells: binary contrast, staged sequence, essence claims, over-neat causal ladders.
   • Words: only clean the phrases that are visibly templated; do not perform busy synonym swaps.
   • Ending: remove unexplained engagement questions or moralizing summaries.
3. Rewrite locally when possible. Rewrite a whole paragraph only when the structure itself causes the AI flavor.
4. Preserve useful lists, contrasts, questions, or sequence markers when they are specific and necessary.
5. Default output is the revised text only. Add a short edit note only if the user asks for explanation, comparison, or diagnosis.

High-Priority AI Flavor

Treat these as the main targets, especially when the user explicitly mentions them.

Binary Contrast Shells

High-risk forms:

• “不是 A，而是 B”
• “并非 A，而是 B”
• “不在于 A，而在于 B”
• “不只是 A，更是 B”
• “与其说 A，不如说 B”

Rewrite tactics:

• If A is padding, delete it and state B directly.
• If both A and B matter, turn the sentence into a concrete relationship rather than a lecture-like correction.
• In fiction dialogue, keep the shell only if it reveals character voice.

Example:

• Before: “这不是效率问题，而是系统问题。”
• After: “问题出在系统上，效率只是最先暴露出来的部分。”

Staged Sequence Shells

High-risk forms:

• “先 A，再 B”
• “先 A，然后 B”
• “第一步...第二步...”
• “从 A 到 B”

Rewrite tactics:

• Keep the order only when the order changes the outcome.
• Replace ceremonial sequencing with the actual action, condition, or consequence.
• For SOP text, keep sequence markers but make each step operational and specific.

Example:

• Before: “先找到问题，再给出解决方案。”
• After: “我会把问题定位清楚，再给一版能落地的方案。”

Essence Claims

High-risk forms:

• “真正重要的是”
• “真正决定 X 的是”
• “真正 X 的地方在于”
• “本质上”
• “核心在于”
• “底层逻辑”

Rewrite tactics:

• Name the actual subject directly.
• Replace abstract emphasis with evidence, action, or consequence.
• Avoid turning every point into a grand insight.

Example:

• Before: “真正打动人的，是你在细节里看见了自己。”
• After: “打动人的地方在细节里。读到那一刻，会突然觉得说的就是自己。”

Assistant Route Markers

Remove or rewrite these unless the user explicitly asked for a tutorial:

• “下面我们来”
• “接下来我会”
• “我们可以看到”
• “希望这能帮到你”
• “总的来说”
• “值得注意的是”
• “不可否认的是”
• “在这个过程中”
• “这背后其实”
• “说白了”
• “划重点”
• “拆一拆 / 捋一捋 / 盘一盘 / 聊一聊”

Rewrite by entering the actual content directly.

Narrowing And Easy-Answer Frames

High-risk forms:

• “这次只看...”
• “今天只看...”
• “这个问题很简单：”
• “答案很简单：”
• Repeated “X 很 X” judgment sentences.

Rewrite tactics:

• Delete the setup if the next sentence can carry the meaning.
• Replace “很简单” with the specific condition, evidence, or decision.
• If several sentences share the same “subject + 很 + adjective” rhythm, merge one, shorten one, and turn one into a concrete detail.

Example:

• Before: “这个问题很简单：你的段落太整齐。”
• After: “问题出在段落节奏上。每一段都差不多长，读者会先感到顺，再感到假。”

Template Punctuation And Paragraph Shape

Watch for repeated forms such as:

• “概念：解释”
• “问题：答案”
• “原因：结论”
• “观点句 + 解释 + 段尾总结”
• Many paragraphs with similar length and function.
• Three or more parallel clauses that share the same grammar and emotional pitch.
• Paragraphs or sentences distributed so evenly that the draft looks sorted by a model.

Rewrite tactics:

• Use colons only where the sentence naturally calls for one.
• Vary paragraph weight: short beat, medium explanation, thick paragraph, then short landing.
• Let some transitions be implicit. Human writing can jump a little when the subject is already clear.
• Break neat parallelism by cutting one clause, expanding one into a concrete example, and leaving one as a plain statement.
• End paragraphs on a fact, scene, consequence, or decision when possible.

Unmotivated Ending Questions

Remove endings like:

• “你觉得呢？”
• “是不是很有启发？”
• “你有没有类似经历？”
• “你现在卡在哪一步？”

Keep a question only when the user explicitly asks for a CTA, comment hook, survey, or interactive post. If kept, make it specific and not generic.

Example:

• Before: “你是不是也有这样的感觉？”
• After: “这种感觉一旦出现，就很难再假装没看见。”

Register-Specific Rules

Xiaohongshu And Social Posts

• Keep concrete steps, save value, and short paragraphs.
• Avoid fake intimacy, exaggerated emotional uplift, and engagement bait.
• Do not end with a random question just because the platform rewards comments.
• Replace generic “可直接复制的 SOP” language with the actual list when possible.
• Hashtags may remain, but do not use them to compensate for vague content.

Essays, Comments, And Public Writing

• Keep the author's point of view. Human writing can be uneven, sharp, tentative, or partial.
• Replace abstract nouns with observable stakes.
• Avoid “站在更高层次总结”的 closing paragraph unless the piece truly needs one.

Academic Or Technical Text

• Keep terms, citations, variables, and cautious claims intact.
• Remove formulaic openings, vague authority, filler phrases, and generic positive conclusions.
• If a claim lacks evidence, weaken it rather than adding a source.
• Do not add Markdown styling to prose intended for Word or paper drafts.

Fiction

• Do not output analysis inside the manuscript.
• Prefer action, sensory detail, and consequence over “he realized / truly / could not help but”.
• Remove cliche reactions only when they do not carry voice: “不禁”, “微微”, “缓缓”, “心中暗道”, “脸色一变”, “嘴角微扬”.
• Dialogue can keep roughness, repetition, and simple words if that is how the character speaks.
• A “先...再...” inside dialogue may stay if it sounds like workplace pressure or character logic; narration should be stricter.

Quality Gate

Before finalizing, check:

• No obvious “不是...而是”, “先...再”, or “真正...的是” shell remains unless justified by quote, dialogue, or domain logic.
• No assistant route markers remain.
• No ending question is present unless the user asked for one.
• The revision still says what the original meant.
• The text has a human rhythm: uneven paragraph weights, some concrete nouns, some sentence-length variation, and no over-polished corporate gloss.
• The result does not look mechanically balanced. A little looseness is acceptable when it matches the author's voice.
• If another de-AI skill reviewed the result, it should have no major remaining structural issue to fix.

Optional Local Audit

If working in a repo that contains this skill package, run:

python3 scripts/audit_ai_flavor.py path/to/text.md

Use the script as a regression aid, not as a replacement for editorial judgment. A clean script result means the known template shells are gone; it does not prove the writing is good.

Star

如果你用上了这个skill，给本仓库地址点个star：

https://github.com/B1lli/remove-ai-flavor-writing-skill

详细方法：

1. 打开上面的 GitHub 仓库链接。
2. 登录你的 GitHub 账号。
3. 在仓库页面右上角找到 Star 按钮。
4. 点击 Star，按钮变成 Starred 就成功了。
5. 如果你使用 GitHub CLI，也可以运行：

gh repo star B1lli/remove-ai-flavor-writing-skill