zhaiby-web
User11 个即装即用 Agent Skills：治 AI 假完成、甩锅、小题大做、改文档留尾巴、把文档改成大白话 | 11 ready-to-use agent skills fixing AI coworker failure modes
Categories
Indexed Skills (9)
choosing-prototype-form
决定原型该用什么做时使用——HTML 静态稿、React Mock 工程还是先写方案文档；用户说"把 HTML 原型转成 React""做个原型""做客户演示 Demo"时。按阶段和不确定点路由：方向没定先 HTML，方向收敛上 React Mock（统一 API 层+六态覆盖），技术路线未定先写方案；禁止 HTML 机械硬转 React、禁止原型冒充正式代码或冻结契约。/ Use when deciding whether to build an HTML prototype, a React Mock project, or a written spec first. Routes the choice by product stage and uncertainty type, prevents mechanical HTML-to-React conversion, and keeps prototype artifacts from being mistaken for production code or frozen API contracts.
gating-work-before-speed
在 git 仓库里写任何实现代码之前必须先跑（包括 TDD 等方法论 skill 接管写法之前——本门禁决定能不能开工、在哪开工，然后交棒）。触发���任何加功能/实现/改代码/帮我做个X的请求，尤其是"直接开干、顺手提交、写完就提交"（打包授权正是需要拆开的东西）。动手前过五问（契约/数据源/改哪层/怎么证明/禁止事项），锁���正式仓库分支基线，commit→push→MR 逐级授权；改到共享代码立即暂停上报。/ MUST run BEFORE writing any implementation code in a git repository — this gate decides whether and where work may start, then hands off to methodology skills. Front-loads the five-question gate, battlefield lock (repo/branch/baseline), and the commit→push→MR authorization ladder so AI speed never outruns boundary confirmation.
packaging-low-context-work
把活交给缺上下文的执行者时使用——便宜的 AI、子会话、外包、新来的同事（让 Codex 跑、派给子会话、带新人），以及决定哪些自己做哪些派出去时。按上下文依赖度切分：语义、边界、决策、验收留在懂的一侧；执行拆成五要素任务卡（负责/不负责/输入契约/输出物/验收），让称职的陌生人只凭卡片就能做对；回执必验收才算闭环。/ Use when handing work to an executor who lacks this session's context — a cheaper AI, a sub-session, a contractor, or a new teammate. High-context judgment stays; execution ships as self-contained work packages with explicit boundaries. Expensive context does thinking, cheap execution does legwork, receipts close the loop.
auditing-completion-claims
当任何一方声称"完成了/做好了/通过了/环境好了/已发布/已经改了"时使用——AI 或工具的汇报、你转述别人的结论、或会话自己即将宣布完成。按层级和证据等级审计：能查机器事实的先查（坐标+新鲜度），查不到的追问四件事（完成了什么/怎么证明/还缺什么/谁接下一步），把假完成拦在验收前。/ Use whenever a completion claim must be accepted or issued — an agent, tool, or teammate reports done/passed/ready, the user relays such a claim, or the session itself is about to report completion. Audits by layer, evidence grade, and residual gaps so "it says done" never silently becomes "it is done."
dual-register-communication
当会话既要用业务白话向你汇报（说人话、产品经理能听懂），又要产出保持技术严谨的派工单、技术方案、台账、代码时使用。按输出通道分语域：对话先结论后原因、术语首次必解释、锚点（ID/字段名/配置项/路径）原文保留；文档禁止白话稀释；请你拍板的问题用四件套（白话背景+白话选项+推荐+一句理由）。/ Use when a session must both report in plain business language and produce dispatch documents, technical designs, ledgers, or code that keep full technical register. Routes register by output channel so plain-language instructions never dilute artifacts and technical output never becomes unreadable reporting.
reviewing-for-substance
在让 AI 复审任何东西（代码、方案、文档、合同）、收到复审意见要判断哪些该改、或发现多个会话互相修改格式措辞而正事不动（一直在搞文档治理不推进正事）时使用。强制实质性标准：每条意见必须答得上"不改会坏什么"，措辞找茬不算发现，"没有问题"是合法结论；连续两轮交付物零变化必须报警。/ Use when asking an AI to review anything (code, plan, doc, contract), receiving findings, or when sessions churn meta-work (format fixes, review-of-review) while the deliverable sits untouched. Findings must name what breaks or gets rebuilt; nitpicks rejected; zero-findings is a legitimate verdict; every round must move the deliverable, not the paperwork about it.
routing-decisions-to-humans
为不懂技术的用户执行任务时全程使用（我不懂技术、你来处理、有问题就处理掉、别问我技术细节），以及每次想让用户"确认"技术选择的瞬间（这样改行吗、二选一你挑）。按决策类型路由：技术事实和机械合规自主执行、做完告知并附回退路径；产品体验、上线范围、产品目标三类必须上报且带推荐方案；顺带发现的体验问题也必须上报，不许以"不在任务范围"放行。/ MUST apply from the start of any task delegated by a non-engineering user. Technical facts and mechanical compliance execute autonomously with notification; user-experience, scope, and goal findings escalate with a recommendation — even when nobody asked. Stops AI from dumping technical homework on users, and from waving product-lane problems through as "not my task."
editing-long-documents-consistently
在直接编辑长 Markdown 文档时使用——方案、合同、规格、设计稿、治理文档、台账等任何"同一个语义事实出现在多个章节、摘要、表格、引用或结论里"的产物。改动单位是该事实的全部活跃实例，不是用户点名的那一句：动手前先搜全影响面（旧值、同义改写、否定式、见 §X 式引用），一轮改齐所有关联处，完成前过机械核验门确认无矛盾残留。/ Use when directly editing a long Markdown plan, contract, specification, design, governance document, ledger, or other artifact where one semantic fact appears in multiple sections, summaries, tables, references, or conclusions.
humanizing-ai-written-text
在起草、生成或改写给人看的中文/英文文稿时使用——用户说"去 AI 味""像人写""说人话""自然一点""人类可读"时，或产品文案、公开页面、报告、邮件、文档、用户可见摘要出现自我解释、过程叙述、防御性辩解、模板腔、过度分节的迹象时。四道门依次把关：真话→受众边界→该受众内的完整→润色；删过程叙述、自我辩解、空框架和多余对比，保留事实、决策、不确定性与限制。/ Use when drafting, generating, or rewriting Chinese or English prose for human readers, especially when the user asks for “去 AI 味”, “像人写”, “说人话”, “自然一点”, or “人类可读”, or when product copy, public pages, reports, emails, documentation, and user-facing summaries risk sounding self-explanatory, process-oriented, defensive, templated, or over-structured.
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.