deanlu.ai

How to Use Claude Cowork Effectively

8 min read

Source Snapshot

  • Origin: Anthropic Claude Code cost docs, Anthropic prompt caching docs, and token-optimization articles
  • Type: Web article and documentation synthesis
  • Author / org: Anthropic, MindStudio, Agensi, and Finout
  • One-line takeaway: Use Claude like a high-leverage coworker: scoped outcomes, clean context, clear checkpoints, and explicit cost guardrails.

Garden Card

This note is a Quartz-ready operating guide for using Claude Cowork or Claude Code without losing cost control, context quality, or review discipline.

这篇笔记是一份面向 Quartz 的 Claude Cowork / Claude Code 使用指南，重点是控制成本、上下文质量和复核纪律。

  • Core question: How should a human delegate substantial work to Claude without creating runaway context or quality drift? 核心问题：人应该如何把复杂工作交给 Claude，同时避免上下文失控或质量漂移？

  • Operational value: It improves delegation quality, reduces token waste, and keeps review at meaningful decision points. 运营价值：它提升委托质量、减少 token 浪费，并把复核放在真正关键的决策点。

  • Best connection: Agent SDK Core Concepts, Introduction to Claude Cowork, SDK Demo Learning Map 最适合连接的内容：SDK 核心概念、Cowork 入门和 Demo 学习路线。

1. Executive Summary

Claude performs best when the task has a clear deliverable, a bounded working area, and a known definition of done. Context management is therefore both a cost control practice and a reasoning quality practice.

当任务有明确交付物、限定工作范围和清晰完成标准时，Claude 表现最好。因此，上下文管理既是成本控制，也是推理质量管理。

The effective pattern is to scope the task, prepare the workspace, ask for a plan when risk is high, execute in checkpoints, verify the output, and clear or compact context when the work changes.

有效模式是先限定任务、准备工作区，高风险时先要计划，再分阶段执行、验证输出，并在工作切换时清空或压缩上下文。

  • Main idea: Context is an operating budget. 主要观点：上下文是一种运营预算。

  • Why now: Agentic work increasingly spans files, browsers, apps, calendars, and codebases. 为什么现在重要：智能体工作越来越多地跨越文件、浏览器、应用、日历和代码库。

  • Where it applies: Website work, Obsidian publishing, Feishu automations, GitHub/Vercel releases, and manufacturing-agent planning. 可以应用的场景：网站工作、Obsidian 发布、飞书自动化、GitHub/Vercel 发布和制造业智能体规划。

Decision Signal

Claude is most useful when managed like a capable colleague: scope the assignment, prepare the workspace, review checkpoints, and clear context when work changes.

2. Key Technical Terms

Use these terms to manage Claude delegation as a controlled operating process.

这些术语可以帮助把 Claude 委托工作管理成受控运营流程。

  • Context window / 上下文窗口: The active working memory used by the model during a task.

    模型在任务中使用的活动工作记忆。

  • Context rot / 上下文腐化: Quality decline caused by accumulated stale history and noisy outputs.

    过期历史和噪声输出累积导致的质量下降。

  • Plan mode / 计划模式: A checkpoint where Claude explores and proposes before acting.

    Claude 在执行前先探索并提出方案的检查点。

  • Prompt caching / 提示缓存: Mechanism for reducing repeated prompt cost when stable context is reused.

    复用稳定上下文时减少重复提示成本的机制。

  • Usage monitoring / 用量监控: Tracking token, cost, latency, and agent fan-out behavior.

    跟踪 token、成本、延迟和智能体扩散行为。

3. Core Notes

3.1 Problem

Unscoped Claude sessions can read too much, run too long, and carry stale context into unrelated work.

没有限定范围的 Claude 会话可能读取过多内容、运行过久，并把过期上下文带入无关工作。

  • Long sessions increase cost and reasoning noise. 长会话会增加成本和推理噪声。

  • Too many connected tools create hidden overhead. 连接过多工具会产生隐藏开销。

  • Vague deliverables cause rework. 交付物不明确会导致返工。

3.2 Mechanism

A good Cowork task names the deliverable, inputs, scope, review boundary, and definition of done before execution begins.

好的 Cowork 任务会在执行前说明交付物、输入、范围、复核边界和完成标准。

  • Use plan mode for complex or risky work. 复杂或高风险工作先用计划模式。

  • Use focused folders and relevant sources only. 只使用聚焦文件夹和相关来源。

  • Use checkpoints for verification and direction changes. 用检查点进行验证和方向调整。

3.3 Evidence

The source set emphasizes /usage, /clear, focused compaction, concise project instructions, prompt caching, and budget awareness.

来源材料强调 /usage、/clear、聚焦压缩、精简项目说明、提示缓存和预算意识。

  • Planning reduces runaway-loop risk. 计划可以降低失控循环风险。

  • Persistent instructions should be lean. 持久指令应该保持精简。

  • Skills are better than repeatedly re-explaining a process. 技能比反复解释同一流程更适合复用。

3.4 Boundary

Claude should execute mechanical work and expose evidence, while the human reserves judgment for facts, business direction, and irreversible actions.

Claude 应执行机械性工作并暴露证据，而人保留对事实、业务方向和不可逆操作的判断。

  • Do not leave long-running agent work unattended. 不要让长时间运行的智能体工作无人看管。

  • Do not let stale context become hidden policy. 不要让过期上下文变成隐藏策略。

  • Do not connect more data than the task requires. 不要连接超过任务所需的数据。

4. Concept Map

Use wikilinks to connect this note into the broader Quartz graph.

使用双向链接把这篇笔记接入更大的 Quartz 知识网络。

flowchart LR
  A["Define Deliverable"] --> B["Limit Context"]
  B --> C["Plan"]
  C --> D["Execute Checkpoints"]
  D --> E["Verify Output"]
  E --> F["Compact or Clear"]
  F --> G["Capture Reusable Skill"]

Diagram labels stay in English for rendering consistency and easier reuse across published pages.

图中的标签保持英文，便于 Quartz 渲染后跨页面复用，也方便技术读者快速识别。

5. My Take

Claude should be managed as delegated work, not casual chat. The more important the task, the more explicit the scope, evidence, approval, and context lifecycle should be.

Claude 应被当成委托工作来管理，而不是随意聊天。任务越重要，范围、证据、审批和上下文生命周期就越需要明确。

  • What changed my thinking: Token management is also reasoning-quality management. 改变我理解的地方：token 管理也是推理质量管理。

  • What I may do next: Convert repeated work into skills instead of re-prompting from scratch. 下一步可能行动：把重复工作做成技能，而不是每次重新提示。

  • What still needs verification: Connected tool scope, prompt caching behavior, and team usage monitoring. 仍需要验证的内容：连接工具范围、提示缓存行为和团队用量监控。

Reuse Path

Convert this note into a Claude task preflight checklist for large work sessions.

References

Graph View

Backlinks