Set up an end-to-end test suite in any repo, following practices that make e2e a reliable per-PR gate: real flows over bypass, layered assertions, a reusable auth/session helper, video+trace evidence, and a compounding suite. Use when a repo has no e2e (or weak e2e) and you want system-level tests — "set up e2e", "add end-to-end tests", "scaffold a test gate".

Scaffold a one-command `dev-local` launcher for ANY codebase. Investigates the repo to find its services, ports, and infra dependencies, then generates a single `scripts/dev-local.sh` (up/down/status/logs/restart) that runs every dev server in one tmux session, plus a short skill doc describing it. Use when someone says "set up dev-local", "make a one-command dev launcher", "I want one script to start this repo", "scaffold dev-local for this project".

Scaffold an isolated CLOUD dev box per agent (via crabbox + Daytona) for any codebase — the parallel-safe counterpart to dev-local-setup. Each agent gets its own full stack (own DB + dev server) and an in-box browser for e2e, so concurrent loops never collide on ports/state. Sets up the snapshot image, .crabbox.yaml, an idempotent setup.sh (also boots the stack locally), and a cbx.sh wrapper. Use when the user says "set up crabbox", "give each agent its own box", "add cloud testing", "make this repo testable in the cloud / on Daytona", "parallel-test this", or when setup-codebase-harness needs true per-agent isolation.

Audit a repo's agent context — CLAUDE.md files, codebase docs, skills, and tool/MCP designs — against Anthropic's Claude 5 context-engineering guidance ("unhobbling": Anthropic cut ~80% of Claude Code's system prompt with no eval loss). Finds overconstraint, conflicting instructions, redundancy, stale facts, and missing "unknown knowns"; produces a scored findings report with concrete rewrites, then applies approved fixes. Use when someone says "audit my CLAUDE.md", "context audit", "unhobble this repo", "review our agent docs/skills/tools", or after upgrading to Claude 5-generation models.

Guide for creating effective skills. This skill should be used when users want to create a new skill (or update an existing skill) that extends Claude's capabilities with specialized knowledge, workflows, or tool integrations.

IMMEDIATELY USE THIS SKILL before answering any task - internal design thinking that identifies the question type, selects applicable skills, and converges on the best approach before implementation

Use when user asks to "read paper", "analyze paper", "summarize paper", "读论文", "分析文献", "帮我看一下这篇paper", "论文笔记", or provides a PDF file that appears to be an academic paper. Specialized for CV/DL papers. Also supports Zotero integration: "读一下这篇论文 ...", "快速看一下这篇论文 ...", "批判性分析这篇论文 ...", "读一下 Zotero 里的 XXX", "批量读一下 Zotero 里 VLA 分类下的论文" **重要触发词**: "读一下 XXX"、"读一下这篇"、"帮我读" → 必须调用此 skill

重新生成 Obsidian 里的目录页 / 导航页（MOC）。 当用户说“更新索引”“更新论文和概念目录”“刷新论文和概念目录”“刷新MOC”时使用。

每日论文推荐的一句话总入口。用户说“今日论文推荐”“过去3天论文推荐”“过去一周论文推荐” “最近3天论文”“看看这周有啥论文”时使用。 内部会自动串联论文抓取、推荐生成、重点论文笔记三步，无需用户手动拆开。

论文点评（3 步流水线的第 2 步）。读取富化后的论文数据，扫描笔记库，生成有态度的推荐点评， 保存推荐文件到 Obsidian，更新 history；git 自动化默认关闭。 触发词："论文点评"、"跑一下论文点评"

论文抓取（3 步流水线的第 1 步）。抓取 arXiv + HuggingFace 最新论文，打分筛选，富化信息， 输出到 /tmp/daily_papers_enriched.json 供后续 skill 使用。 触发词："论文抓取"、"跑一下论文抓取" 支持多天模式："过去3天论文推荐"、"过去一周论文推荐"、"过去一周的论文"、"抓 3 天的论文"、"最近5天"

Deprecated 兼容入口：旧 cs-roadmap 调用用；转交 cs-epic --stage planning。新请求不要主动选择。

Deprecated 兼容入口：旧 cs-roadmap-impl-goal 调用用；转交 cs-epic --stage goal-package。新请求不要主动选择。

能力愿景维护。触发：起草/回填/更新 requirements，或 design/acceptance 需要 req delta。不要用于单个功能实现方案(cs-feat design)、大需求拆解排期(cs-epic)、开放式发散讨论(cs-brainstorm)、领域术语/ADR建模(cs-domain)。

Deprecated 兼容入口：旧 cs-refactor-ff 调用用；转交 cs-refactor --mode fastforward。新请求不要主动选择。

attention 短规则追加。触发：一两行、长期有效、每次会话都必须知道的项目规则；详细经验走 cs-keep，结构性决策走 cs-domain。

Deprecated 兼容入口：旧 cs-issue-fix 调用用；转交 cs-issue --stage fix。新请求不要主动选择。

Deprecated 兼容入口：旧 cs-issue-analyze 调用用；转交 cs-issue --stage analyze。新请求不要主动选择。

Goal 自主达成。触发：明确终点/验收/预算、持续迭代到完成，或 grill me 后开工。不要用于单个功能的规划与实现(cs-feat)、大需求端到端拆解(cs-epic)、bug 诊断修复(cs-issue)、行为等价重构(cs-refactor)、对外文档(cs-docs)；也不用于没有实现终点的纯 design / roadmap / brainstorm / audit。

CodeStable 使用反馈闭环。仅在用户明确要求记录或上报 cs skill 跑偏、工具失败、规则不清、agent 被纠正时使用；采集本机证据并准备可确认的公开 issue。