AI 工作流工具集

针对单个项目，根据用户提供的上下文生成带优先级的计划初稿，确认后产出周/月维度的执行计划。 适合有明确目标和时间约束的单项目规划。不处理多项目汇总，那是 daily-schedule 的职责。

当用户说"token 消耗太快"、"为什么这么贵"、"比之前贵了好多"、"诊断 token"、"检查配置"、"token audit" 时使用。 诊断 Claude Code 环境中隐藏的 token 消耗热点，输出按影响程度排序的优化建议。 不要用于：分析单次对话的 token 用量统计；不要修改任何配置（纯只读诊断）。

Use when the user wants to create a Claude Code hook, 钩子, 自动触发, 拦截工具调用, 每次AI停止时, 看门狗, 文件保护, 任务通知 - guides through an interactive interview to gather trigger event, handler type, scope, and generates both the settings.local.json config and hook script. Covers: task notification, file protection, watchdog, file header, remote monitoring, and custom hooks.

Applies 12 meta-reflection techniques as active execution checkpoints for complex, high-reliability tasks. MANUAL triggers: "元反思", "AI鲁棒性", "任务风险", "让AI更可靠", "meta-reflection", "enhance robustness". AUTO triggers (no user invocation needed): - Task mentions 5+ modules/systems/entities - Keywords detected: "生产环境", "production", "数据迁移", "关键业务", "不能出错", "zero downtime" - Another skill completes a complex task and requests post-execution reflection DO NOT invoke: simple queries, mid-execution interruption, user asks only for code/answer.