本文由 AI 分析生成
建立時間： 2026-03-03 來源： https://x.com/KKaWSB/status/2028677808713822310
Summary
A beginner masterclass on Claude Cowork aimed at non-technical users, focused on building a reusable collaboration system rather than memorizing steps. It distinguishes Chat (conversation), Code (development), and Cowork (delegation), then walks through workspace setup, three core context files (about-me, voice-style, working-rules), scheduled tasks, sub-agents, connectors, and reusable skill files.
這是一堂面向非技術用戶的 Claude Cowork 新手大師班，重點是建立可復用的協作系統，而非死記操作步驟。文章區分 Chat（對話）、Code（開發）、Cowork（委派），接著講解工作空間搭建、三個核心上下文文件（about-me、voice-style、working-rules）、定時任務、子智能體、連接器，以及可複用的技能文件。
Key Points
- Cowork is delegation, not chat: you describe what “done” looks like and walk away while it plans, executes, and delivers.
- Never authorize the whole Documents folder — create a dedicated ~/Claude-Workspace/ with _context, projects, templates, outputs, archive.
- Three context files make Claude “know you”: about-me.md (identity/priorities), voice-style.md (writing voice, with real samples), working-rules.md (how to collaborate, what’s forbidden).
- Auto-load context via global instructions or a per-project _MANIFEST.md with read-tiers (must-read / on-demand / never-read).
- A good task answers three questions: what does done look like, what are the constraints, what to do when uncertain.
- Scheduled tasks (/schedule) automate weekly reports and competitor monitoring; sub-agents parallelize comparisons; skill files capture repeated complex workflows.
Insights
The thesis — “ChatGPT era chased smarter prompts; Cowork era chases better systems” — reframes prompting from a per-use cost to a one-time infrastructure investment. The author argues 90% of unstable output comes from poorly written context files, so the leverage is in the system (folders, context, manifests, skills), not in any single prompt. This mirrors the CLAUDE.md philosophy applied to non-coding knowledge work.
Connections
Raw Excerpt
提示词是一次性的——每次都要重新写。系统是可复用的——搭建一次，持续受益。投入2小时搭建系统，每天都能赚回来。