yousa-skills

Use when the user asks to correct English prompts, broken English, typos, grammar, wording, tone, workplace English, business writing, presentation language, or asks for English practice, quizzes, or help expressing ideas clearly in English.

Use when 用户要求提取、提炼、总结某个日期或时间范围的群聊/聊天记录（EACC 投资群 / eacc-chat），生成群聊日报、核心话题分析或投资价值信息筛选。触发词：提取聊天记录、群聊提炼、群聊总结、某月某日群里聊了什么、chat digest、群聊日报、群聊干货。

评估或对比会议纪要/访谈整理/深度纪要的质量：对照原始转录核查保真度，盘点覆盖度，输出多维度打分与失真清单。适用时机：(1) 用户问"这两份纪要哪份质量更高/写得更好"，(2) AI 生成的纪要发布前验收（有没有幻觉、编造、立场写错），(3) 给纪要挑毛病/做质检/打分。触发语：评估纪要质量、对比纪要、纪要靠谱吗、纪要验收、纪要打分、哪份纪要好。

对会议纪要/活动回顾/访谈整理等将对外发布的文档做敏感信息检查与脱敏，覆盖查塔姆宫规则（Chatham House Rule）合规：人名机构名及变体、职级暗示、玩笑话与人事信号、点评同行、口语化敏感词、外链击穿匿名、身份指纹拼图、第三方未公开信息、编辑残留。适用时机：(1) 纪要将对外发布/进知识库/发群前，(2) 嘉宾或机构方要求脱敏审稿，(3) 用户提到查塔姆规则/讲事不讲人/匿名化/敏感信息检查。输出违规清单+修复建议，或直接执行脱敏。

Use when designing or reviewing lifecycle status fields, finite state machines, async task/order/payment/upload/approval workflows, or when states are overlapping, action-shaped, ownerless, unsafe under retries, concurrency, timeouts, or terminal-state updates.

深度研究与验证技能：当用户提出需要多源信息搜集、逻辑推理和事实验证的复杂问题时使用。 适用场景包括但不限于：投资研究、行业分析、技术对比、政策解读、商业决策支持等需要 "先想清楚怎么查，再去查，查完还要验证"的场景。 触发关键词/短语：为什么A而不是B、分析一下、帮我研究、设计一个SOP流程来回答、 深度分析、请验证、fact check、多角度分析、请先理解问题再回答、不要立即回答。 当用户明确要求"先设计研究流程再回答"或"验证你的结论"时，务必使用此技能。 当用户的问题涉及多个实体的比较（公司、技术、政策等）且答案不是简单事实时，也应触发。

对一段陈述/视频/文章/分析报告做真实性验收：拆解论点链 → 联网核查关键数据 → 找出反方证据 → 输出可点击的 8 维度红绿灯诊断与最终客观判断。当用户说"帮我看看这个说法靠不靠谱 / 验证真实性 / 反方观点是什么 / 这是不是泡沫 / 帮我做一个验收 / 这观点合理吗"时调用。

对金融/宏观/政经/商业叙事做真实性与逻辑性校验：拆命题→查数据→验结构→找反例→打分→出验收结论。当用户说"帮我验证下这段话 / 这个说法靠谱吗 / 找反例 / 是不是叙事过拟合 / 真伪核查 / 二次考察"时调用。

"把分析结果 + 联网检索证据整合成可长期归档的中文知识笔记（带元信息、证据表、方法论、复现方法、AC 自检表）。当用户说"整理成知识笔记 / 归档 / 整理成 Obsidian 笔记 / 沉淀为档案"时调用。"

"对 YouTube 视频做四步深度分析：抓字幕 → 结构化总结 → 联网事实核查 → 输出"合理/不合理/启发"判断报告。当用户给出 YouTube 链接并要求分析、梳理、批判性解读、判断真伪、提炼观点时调用。"

将会议/圆桌/访谈的录音转录文本（含说话人标签）转化为结构化深度纪要——按「判断 → 思维模型(mindset) → 实证案例」三层提炼，聚合核心议题，联网核证业界实践，多 agent 并行编排并验收。适用于话题分散、多人实时交互、传统时间线式 AI 纪要抓不住重点的对话（圆桌、闭门交流、嘉宾问答、podcast）。触发语：深度纪要、会议纪要总结平庸/没抓住重点、转录转纪要、圆桌整理、从转录中识别核心嘉宾并整理纪要。

Review or write Python code with a 3AM debuggability lens, focusing on readable tracebacks, visible async/background ownership, exception semantics, hidden side effects, mutable state capture, callback/closure chains, dynamic magic, and low-value indirection.

评审或新写 Go 代码时，用“凌晨 3 点能 debug”原则检查 panic stack 可读性、异步边界可见性、mock-only 间接层、callback 参数、闭包链、隐藏 goroutine 和单 caller 包装。

Use when a research topic has unclear scope, ambiguous terms, missing evidence criteria, or is likely to become a source-code learning study, investigation, audit, architecture explanation, or design inquiry

Use when identifying unused functions, dead code, deletion candidates, or "safe to remove" symbols, especially with RPC, config, generated code, scripts, or exported APIs

Use when judging whether an activity, skill, project, workflow, career choice, learning path, or repeated effort has compounding value, durable payoff, 复利属性, or is mostly one-off consumption.

Use when drafting, reviewing, or updating pull request descriptions for open source projects where maintainers need issue context, behavior changes, tests, tradeoffs, or CI status to review confidently.

Use when asked to help learn a specific feature in a codebase and produce a guided self-study document or SOP. Triggered by requests like "帮我设计一个学习源码的文档", "如何系统学习这段代码", "生成一份源码阅读指南", or when the goal is to create a reproducible learning workflow for a repository feature.

Use when starting implementation, bugfix, refactor, docs, or investigation work in a repository where upstream pull requests may already cover or conflict with the task.

Use when creating docs/todo entries, follow-up work records, or post-validation TODOs that future humans or agents must understand without chat history

Use when creating local git commits in codetok from existing changes, especially when scope, staging, verification, or message clarity could drift.

Use when implementation is already complete, a GitHub PR exists or must be opened, and follow-up work is still needed because Copilot or other AI review comments may arrive after the initial push.

Use when an article compares multiple entities (companies, approaches, strategies) in parallel and you suspect the comparison may be artificial, unequal, or evidence-driven by narrative rather than fact. Symptoms include "A took route X, B took route Y, C took route Z" structures, one entity credited with a feature others quietly share, comparing a specific tactic against a strategic principle, or conclusions that exist to make the story prettier.

Use when submitting a skill (or a directory of skills) to the yousa-skills GitHub project. Handles copying skill files, updating docs/readme/skills.json, regenerating README.md and README.zh-CN.md, committing changes, and creating a pull request.

Use when user has an open PR and wants to automatically fix AI code review comments (Copilot, Codex, etc.) via local cron. Triggered by "自动修复 review", "盯着这个 PR", "auto-fix PR reviews", "帮我盯下这个 PR 的 review".

Use when about to git commit, git push, or create PRs. Scans staged/changed files for leaked secrets (API keys, tokens, passwords, credentials). MUST run before any git commit or push operation.

Use when proofreading Chinese articles for typos, grammar issues, and CJK-Latin formatting problems. Triggered by requests like "校对", "查错别字", "语病扫描", "typo检查", "帮我检查下文章". Also use when user pastes Chinese text and asks to "看看有没有问题".

Fetch authenticated or JS-rendered web pages via Chrome CDP (DevTools Protocol) and convert them to clean Markdown files. Use when: (1) user provides a URL that requires authentication (e.g. claude.ai, chatgpt.com, Google Docs, Notion, Confluence, Jira), (2) user wants to save/archive a web page as Markdown, (3) WebFetch fails with 403/401 or returns empty content, (4) the page is a JS-rendered SPA that needs a real browser. Trigger words: 'fetch page', 'grab page', 'save as markdown', 'convert to md', 'pull down this page', 'CDP', 'archive this URL'.

双镜例证法：为文章论点寻找高洞见力的例子。当用户需要为某个主题/观点找有说服力的例证时使用。产出两类例子：(1) 远镜/A类——跨越几百年甚至几千年的组织级案例，证明人性与结构不变；(2) 近镜/B类——身边司空见惯但点破即顿悟的例子。核心是「先找机制，再找故事」，通过同构机制而非相似故事来挖掘深度。

Use when a task is complete and the user asks what was done, how it was done, why those choices were made, or wants a Feynman-style walkthrough of the key concepts

首席社交策略顾问：分析对话记录/聊天截图中的潜台词、情绪和权力动态，提供多维度回复策略。当用户需要：(1) 分析一段对话或聊天记录的深层含义，(2) 不知道怎么回复某条消息，(3) 想理解对方的真实意图，(4) 需要高情商/设立边界/幽默/直击痛点的回复建议，(5) 贴出聊天截图或对话文本请求分析时使用。触发关键词：分析对话、怎么回复、帮我想想怎么说、社交策略、对方什么意思、聊天分析。

通过 tmux 编排多个 Claude Code 进程来批量处理同质任务。当前 CC 作为调度方， tmux 中的 CC 作为执行方。支持 work-stealing 并发调度、任务内建质量门禁。 当用户需要：(1) 批量处理多个独立同质任务（论文分析、代码审查、文档翻译等）， (2) 单次 prompt 处理质量不够需要拆分独立执行，(3) 需要并发加速批量任务时使用。 触发词：批量处理、tmux dispatch、并发执行、逐个处理。