来自 cloud code 团队的 Tharic 解释了如何使用 /loop、/goal 和工作流让 Claude 长时间自主运行，并演示了视频编辑。
Tharic from the cloud code team explains how to use /loop, /goal, and workflows to make Claude run autonomously for long periods, with a demo of video editing.
要点 · TL;DR
循环和工作流让 AI 智能体能够长时间自主工作。 Loops and workflows let AI agents work autonomously for long tasks.
迭代探索规划是与人类的持续对话，而非静态文档。 Planning is an iterative exploration with the human, not a static document.
简洁、基于原则的提示词通常优于详细约束。 Simpler, principle-based prompts often outperform detailed constraints.
核心观点 · Key points
循环、目标和工作流使智能体能够长时间自主运行。 Loops, goals, and workflows enable agents to operate autonomously for extended periods.
规划是迭代探索未知的过程，而非一次性的静态文档。 Planning is iterative exploration to uncover unknowns, not a single static document.
更智能的模型需要更少的指令和示例；精简的系统提示可提升性能。 Smarter models require fewer instructions and examples; leaner system prompts boost performance.
独立的验证智能体可减少自我引用偏差，产生更高质量的结果。 Separate verification agents reduce self-referential bias and yield higher-quality results.
技术成长源于理解系统约束和权衡，而不仅仅是语法。 Technical growth comes from understanding system constraints and trade-offs, not just syntax.
反共识 · Contrarian takes
规划不是一次性规范，而是人机之间持续的对话。 Planning isn't a one-time spec; it's an ongoing human-agent dialogue.
简短、基于原则的提示通常比详细的规则和示例更有效。 Shorter, principle-driven prompts often work better than detailed rules and examples.
“智能体即员工”的比喻有用但有限，智能体记忆和能力不均。 The 'agent as employee' metaphor is useful but limiting; agents have uneven memory and capabilities.
过多约束会阻碍性能；提供理由比绝对禁止更有效。 Excessive constraints can hinder performance; providing reasons is more effective than absolute prohibitions.
智能体会输出冗长计划，人类容易略读，但仔细阅读对避免错误至关重要。 Agents output lengthy plans that humans tend to skim, but careful reading is crucial to avoid mistakes.
本期章节 · Chapters（共 30）
循环介绍与概念Introduction and Concept of Loops
团队使用与任务依赖Team Usage and Task Dependency
详细规划的重要性Importance of Detailed Planning
演示：视频编辑流程Demo: Video Editing Workflow
视频编辑工作流介绍Introduction to video editing workflow
赞助商：Whisper FlowSponsor: Whisper Flow
规划流程与迭代探索The planning process and iterative exploration
理解未知与探索性规划Understanding unknowns and plan as exploration
视频识别的迭代过程Iterative process for video recognition
团队使用HTML产物Team use of HTML artifacts
覆盖样式生成与迭代Overlay style generation and iteration
人机规格演进Specs evolution for human and agent readers
构建低成本原型并迭代Building cheap prototypes and iterating
Claude在Slack与终端的分配How Claude usage is split between Slack and terminal
多任务与Slack集成ClaudeMultitasking and Slack Integration with Claude
代理即员工隐喻Agent as an Employee Metaphor
展示播客制作技巧Showcasing a Podcast Production Skill
技能与工作区Skills and Workspaces
缩略图与图像生成Thumbnails and Image Generation
工作流与子代理Workflows and Sub-Agents
上下文管理技巧Context Management Tips
专注于单一项目Focus on single project
代码质量担忧Concerns about code quality
精简上下文与提示Trimming context and prompts
原则优于约束Principles over constraints
变得更加技术化Becoming more technical
学习技术的原因Reason for learning technical skills
通过HTML报告深度学习Deep learning through HTML reports
Claude Code协作未来Future of Claude Code collaboration