Linuxdo Application

SKILL内容创作
内容创作 未标注权限 许可证不明确
安装这个技能

为 Linux.do 注册撰写高通过率、去 AI 化的中文小作文。进行自适应调查以了解申请人的真实背景,...

累计安装 未标注 本周 未标注
GitHub 星标 未标注 仓库 未标注
下载量 未标注 技能市场下载数
权限级别 未标注权限 安装前请确认风险范围
许可证 许可证不明确 未说明

安装

ClawHub 市场
在支持 Skills 的客户端里执行,装完问一句「你有哪些技能」验证
skillhub install clawhub_clarezoe/linuxdo-application

技能档案

内容创作
字段来自技能仓库与公开索引,中文说明由本站标注
作者 / 维护方clawhub_clarezoe仓库 未标注 · 版本 未标注
宿主客户端未标注技能可被这些客户端加载
安装方式ClawHub 市场安装包 clawhub
权限级别未标注权限安装前请确认风险范围
许可证未说明许可证不明确,商用前请自行确认条款
仓库可用性available技能仓库当前是否可访问
最近核验未标注上游最后核验时间 · 数据更新 1791710504
正文体量6.2 KBSKILL.md 原文大小,越大说明说明书写得越细

SKILL.md 全文

6.2 KB
模型实际读到的说明书原文,来自技能仓库

Linux.do Application Writer

Generate a natural, high-pass-rate registration application for Linux.do through
an adaptive interview followed by rule-aware, de-AI'd text generation.

Triggers

  • "write linux.do application"
  • "linux.do 小作文"
  • "linux.do 注册申请"
  • "help me apply to linux.do"

Quick Reference

Input Output Duration
Answers to 5-8 survey questions Plain text Chinese application (~80-150 chars) 5-10 min

Process

Phase 1: Applicant Survey

Conduct an adaptive interview in Chinese to collect the 4 required info blocks.
Ask one question at a time. Adapt follow-ups based on answers.

Core questions (ask in Chinese, adapt order based on flow):

  1. 你平时主要做什么?(工作、学习、兴趣方向)
  2. 你是怎么知道 Linux.do 的?(搜索、朋友推荐、看到某个帖子?)
  3. 你在上面浏览过哪些内容?有没有印象深的帖子或话题?
  4. 为什么现在想注册?有什么具体的需求或场景吗?
  5. 注册之后你打算怎么用?(潜水、回帖、关注某类话题、分享经验?)

Adaptive rules:
- If an answer is vague (e.g., "想学习交流"), probe deeper: "具体想学什么?在哪看到过相关讨论?"
- If an answer already covers multiple blocks, skip redundant questions
- If the applicant mentions a specific post/topic, ask them to elaborate — this is gold
- Stop when all 4 info blocks are covered (see references/linuxdo-rules.md)

Verification: Before moving to Phase 2, confirm internally:
- [ ] Background covered (what they do/follow)
- [ ] Discovery path covered (how they found the site)
- [ ] Join reason covered (why register now, specific scenario)
- [ ] Usage plan covered (what they'll do with the account)

If any block is missing, ask one more targeted question.

Phase 2: Draft Generation + Risk Check

Generate the application draft using collected info.

Generation rules:
1. Write in first person, casual Chinese — like telling a friend why you signed up
2. Target 80-150 characters. Not too short (looks lazy), not too long (looks try-hard)
3. Weave all 4 info blocks naturally — do NOT use a 4-paragraph structure
4. Use the applicant's own words and phrasing where possible
5. No greetings, no sign-offs, no "你好" or "谢谢" — just the substance
6. Allow sentence fragments, colloquialisms, and imperfect grammar
7. Vary sentence length: mix short punchy lines with longer ones

Risk check — scan the draft against these (see references/linuxdo-rules.md):

Risk Level Check
HIGH Does it mention background? Discovery path? Join reason?
HIGH Is there any concrete fact, or is it all fluff?
HIGH Does it look like a template that could apply to anyone?
MEDIUM Is it only admiration/flattery without substance?
MEDIUM Is it only "I can contribute X" without explaining why here?
MEDIUM Is info density too low for the character count?

If any HIGH risk is triggered, rewrite before proceeding.

Phase 3: De-AI Polish + Final Output

Run the draft through the full de-AI checklist (see references/de-ai-checklist.md).

Audit steps:
1. Check all 12 AI markers — flag and rewrite any matches
2. Read the text as if you're a human reviewer — does it smell like AI?
3. Apply the 3-question smell test:
- Could you swap details and reuse this for someone else? → too generic
- Does every sentence add new information? → cut filler
- Would you text this to a friend? → loosen if too formal

Output: Present the final application as plain text in a code block.
Tell the user in Chinese: "这是你的申请文,可以直接复制粘贴到注册页面。"

If the user wants changes, revise and re-run Phase 3 checks.

Anti-Patterns

Avoid Why Instead
Parallel structures (我喜欢X,我热爱Y) AI marker #1 Vary grammar patterns
三段式 (首先/其次/最后) AI marker #2 Drop scaffolding
AI vocab (赋能/深耕/沉淀/赛道) AI marker #5 Use plain spoken Chinese
Flattery (久仰大名/慕名而来) High-risk per rules Replace with real discovery story
Generic motivation (想学习交流) Medium-risk per rules State specific use case
Press-release tone AI marker #9 Add 吧/嘛/其实/反正

Verification

After final output, confirm:

  • [ ] All 4 info blocks present
  • [ ] 80-150 characters
  • [ ] Zero high-risk patterns
  • [ ] Passes all 12 de-AI markers
  • [ ] Reads like a real person wrote it
  • [ ] Plain text, no formatting

Extension Points

  1. Rule updates: When Linux.do changes registration rules, update references/linuxdo-rules.md
  2. New AI markers: As AI detection evolves, add markers to references/de-ai-checklist.md
  3. Multi-community: Adapt the survey + rule framework for other invite-only communities

References

Related Skills

Skill Use When
humanizer-zh Additional de-AI processing for Chinese text
psychology-master Deeper user profiling during survey
wechat-compliance-check Need to check Chinese content for platform compliance

Always run openskills list before starting to discover and load any relevant skills that can enhance the application writing process.

热门专题

按场景挑技能
全部专题 →

技能内容以作者仓库为准;分类、权限级别与许可证状态的中文说明为中文整理,最近核验 未标注。