Social Text Websense Gate

Purpose

This skill is a text taste gate for social-platform Chinese copy. It does not publish, scrape, score, or render visuals. It helps the content role decide whether a draft sounds like a real internet-native person reacting to a concrete situation, or like an editor packaging a methodology.

The core rule:

网感来自荒谬点识别，不来自网络词密度。

Words like 疑似、绷不住、抽象、这娃 can appear naturally, but they are not the goal. If the copy mechanically stuffs those words without finding the actual absurdity, it still fails this gate.

When To Use

Use this skill when working on:

• Xiaohongshu / Rednote titles, body copy, carousel text, captions, or comment prompts.
• Douyin titles, captions,图文正文,口播开头, or short social text.
• Tech-social posts about AI coding, Agents, Codex, Skills, workflow, GitHub repos, tool failures, non-CS product building, or automation.
• A user says the copy is 太正经, 像编辑, 像 README, 没网感, 像 AI, 像发布会, or 不是经常冲浪的人.
• A draft is factually right but scroll-stopping power is weak.

Do not use it as a replacement for:

• humanizer-zh: general anti-AI-writing cleanup.
• cheat-on-content: scoring, blind prediction, publishing registration, and retro.
• xhs-publish-assistant: final copy-ready publish bundle.
• Fact checking. This skill can make copy sharper, but cannot invent facts.

Default Sequence

For social-platform tech copy, default to this order:

处境 / 情绪 / 自嘲
→ 真实事故感
→ 技术锚点
→ 仓库 / 方法 / 试用 / 反驳入口

Do not open with macro claims like:

AI Agent 工作流正在成为新的生产力范式。
提示词不够用了，Skill 才是资产。
普通人必须掌握 AI 时代的底层逻辑。

Those may be true, but on Xiaohongshu/Douyin they often feel like a technical briefing. Start with the human friction instead.

The Websense Test

Before editing, answer these questions:

1. What is the concrete situation? Who is tired, annoyed, tempted, confused, or being made responsible?
2. What is the absurd point? What exactly is funny, annoying, or screenshot-worthy?
3. What is the technical anchor? Which real tool, behavior, script, workflow, boundary, or failure mode keeps the joke grounded?
4. What is the human relationship? Is the tool helping, over-helping, replacing, exposing, embarrassing, or dumping responsibility back on the person?
5. What action should the reader take? Save, comment, try the repo, ask for the script, argue, or go test it.

If you cannot answer #2, do not add internet words. Find the absurdity first.

Rewrite Patterns

1. Absurd action + plausible annoying excuse

Good social copy often shows an AI/tool doing something over-free, then gives it a painfully reasonable excuse.

Weak:

Codex 有时候会乱改项目。

Stronger:

你让它修一个按钮，它把半个项目翻新，然后跟你说遵循一致性设计。

Transfer pattern:

你让它做 A，它反手做 B，然后给一个听起来很合理但很欠揍的理由。

More examples:

你让它删一段废代码，它顺手整理了整个目录结构，然后说这是降低复杂度。
你让它看一眼报错，它开始讲技术选型的历史沿革。
你让它写个测试，它把测试写成了对业务逻辑的哲学批判。

2. Do not describe the persona; show what the tool did to that person

Weak:

我是一个有很多想法但很累的程序员。

Stronger:

轻松干翻我这种有无数想法、但被上班榨到一行代码都不想写的技术中登。

The second line works because it defines a relationship: the tool has more energy than the person who is supposed to control it.

3. Translate technical mechanics into human situations

Weak:

我们需要为 Agent 增加边界和校验机制。

Stronger:

自由到我想给它办入职、配工牌、安排直属领导。直属领导不是我，是那个校验脚本。

Use the human scene to carry the technical anchor. Do not hide the anchor; bury it in the joke.

4. Keep the repo/method call-to-action dirty enough to feel real

Weak:

欢迎体验我的开源工作流项目。

Stronger:

请大家的大 Codex 狠狠抽打我的小 skill。有用就 star，没用就来骂，我正好修。

The call-to-action should feel like a real builder asking for stress, not a brand account asking for engagement.

Bad Patterns

README tone

本项目基于角色化 Agent 工作流，提供可复用的任务拆解与质量控制能力。

Why it fails: technically clear, socially dead. Move the technical claim after a concrete situation.

Technical launch-event tone

我们正在从 prompt 时代走向 skill 时代。

Why it fails: it starts from trend commentary instead of reader friction.

Meme-word stuffing

Codex 疑似抽象，绷不住了，这娃太离谱。

Why it fails: the words gesture at internet culture but do not reveal a specific absurdity.

Sterilized safety rewrite

需要注意的是，不同工具在不同项目中的表现存在差异，建议结合自身情况谨慎使用。

Why it fails: it may be safe, but it kills the social post. Keep factual boundaries in the background; do not open like a production announcement unless the topic requires it.

Review Checklist

Before returning social-platform text, check:

• The first screen has a human situation, not a macro trend.
• The draft identifies a specific absurdity.
• At least one line could plausibly be screenshot and sent to a group chat.
• The technical anchor is real and not blurred by the joke.
• Internet words are not used as a checklist.
• The copy does not sound like README, launch notes, or a platform announcement.
• The self-mockery lands on a concrete group or state, not generic 我很累.
• The ending gives a clear reader action: try, save, comment, argue, test, or visit the repo.

Output Modes

When asked to revise a draft, use one of these modes:

Gate report

## 结论
通过 / 需要重写 / 只需小修

## 最大问题

## 最该保留

## 重写方向

## 可直接替换句

Rewrite only

If the user asks for direct copy, return only the revised copy. Do not add title pools, hashtags, visual notes, or publishing advice unless asked.

Common Pitfalls

1. Turning websense into a word list. If a prompt says 多用疑似/绷不住/抽象, correct it: those words are optional byproducts, not the target.
2. Polishing away the weirdness. Some roughness is the asset. Do not upgrade every line into a clean methodology sentence.
3. Forgetting the technical anchor. A funny line without the real tool behavior becomes generic meme content.
4. Over-explaining before the hook lands. Social copy earns explanation after it creates recognition.
5. Pretending platform taste is universal. This gate is for social text. Long-form WeChat, docs, README, and investor/legal/medical copy need different tone.