扣子(Coze)Skills+OpenClaw 实战零基础玩转 AI 智能体 (邢云阳 著)（Z-Library）

# 【One-Line Pitch】 A hands-on, zero-code guide for non-technical readers to build reusable AI "Skills" (expert prompt packages) on the Chinese low-code platform Coze (扣子), covering six real-world projects from livestream sales scripts to PPT generation, plus OpenClaw deployment. Read this if you want to turn your personal expertise into an AI-powered assistant without writing a single line of code. # 【Book Arc】 - **Opening (~0%–10%)**: Introduces the book's core mission — bridging the gap between powerful AI agents and non-technical users. Explains why Skills (Anthropic's zero-code "expert experience packages") matter, how they solve the two barriers of workflow orchestration and MCP protocol complexity, and sets up the Coze platform ecosystem (Coze Programming for development, Coze for usage). - **Early (~10%–23%)**: Builds the conceptual foundation. Covers what agents are (from tool-calling to deep-thinking to multi-agent systems), the evolution of agent design patterns (Function Calling, ReAct, planning, reflection, memory), and then dives deep into Skills — their folder structure (SKILL.md, scripts/, references/, assets/), the L1/L2/L3 progressive loading mechanism that prevents "context pollution," and essential Markdown syntax for writing SKILL.md files. - **Early (~23%–32%)**: Transitions to hands-on platform usage. Demonstrates the paradigm shift from low-code visual programming to AI-native natural language programming — you simply describe what you want in conversation, and the platform generates the agent or Skill automatically. Walks through building a weather assistant agent and introduces the Coze product (formerly Coze Space), including its skill store and custom Skill upload. - **Middle (~32%–48%)**: First full project: the Livestream Sales Script Reuse Skill. Shows the complete workflow from requirements analysis (why top streamers' tacit knowledge is hard to transfer), to structuring a reusable script template library (聚人/留客/锁客/说服/催单/引导下单 stages), to writing the prompt, uploading reference documents, and understanding the interactive development sandbox interface. - **Middle (~48%–60%+)**: Continues the livestream project with the actual SKILL.md file structure — YAML front matter, task objectives, step-by-step operation procedures, resource indexing, usage examples, and notes. Establishes the template pattern that subsequent projects (stock analysis, WeChat article images, Xiaohongshu content, PPT assistant, GEO/SEO) will follow. - **Late (~60%–100%)**: Covers the remaining five projects of increasing complexity (text-only → image-text multimodal → multi-Skill collaboration), then moves to OpenClaw deployment and integration with Coze Programming, showing how to add and use Skills in the OpenClaw environment. # 【Key Takeaways】 - **Skills are "expert experience packages" — structured prompts, not code** (Early): A Skill is a folder containing a SKILL.md file (required) plus optional scripts/, references/, and assets/ subfolders. It packages your thinking process, execution steps, tools, and output standards into a reusable format that any agent can read and execute. This is the core mental model for everything in the book. - **The L1/L2/L3 progressive loading mechanism prevents context pollution** (Early): When an agent loads a Skill, it only reads the L1 tag/description first, then loads the full SKILL.md (L2) when a task matches, and finally pulls in scripts/references/assets (L3) only as needed. This solves the "memory bread" problem — irrelevant information crowding out what actually matters in the model's limited context window. - **Natural language programming has replaced low-code as the primary development paradigm** (Early): Instead of dragging nodes and configuring conditional branches, you now describe your business need in conversation — "create a weather assistant that queries weather by time and city" — and the platform auto-generates the agent, tools, and code. The book's core promise: if you can explain the business logic, AI handles the technical details. - **Skills must be deployed to be useful — they're not standalone apps** (Early): Unlike agents that run directly, a Skill is a "package" that needs to be loaded into an agent (like Coze) to take effect. The workflow is: develop in Coze Programming → deploy → call from Coze's skill store or conversation. Understanding this ecosystem architecture is essential before starting any project. - **The livestream script project demonstrates the full Skill development pattern** (Middle): The first project walks through the complete pipeline: analyze why top streamers' expertise is hard to transfer (会做不会教, experience lost with turnover), structure their tacit knowledge into a template library (welcome scripts, engagement hooks, urgency tactics), save as a Markdown reference file, then write a prompt that instructs the agent to generate personalized scripts from product inputs. This becomes the template for all subsequent projects. - **A well-structured SKILL.md follows a consistent framework** (Middle): The generated SKILL.md uses YAML front matter (name + description) followed by: task objectives → operation steps (broken into executable stages) → resource index → notes/caveats → usage examples. This structure ensures the agent knows when to trigger the Skill, how to execute it step-by-step, and what output quality looks like. - **The book's six projects progress from simple to complex deliberately** (Middle): Starting with pure text Skills (livestream scripts, stock analysis), moving to multimodal image-text creation (WeChat article images, Xiaohongshu posts), and finally to multi-Skill collaboration with AI programming (PPT assistant, GEO/SEO). Each project reuses the same development pattern, so by project three you're focusing on business logic, not platform mechanics. # 【Reading Tips】 - **Skim the agent theory sections (Chapter 1) if you're eager to build**: The historical evolution from GPT Store to ReAct to deep-thinking agents is useful context, but the actionable content is the Skills structure (folder components, L1/L2/L3 loading) and the Markdown syntax review. If you're impatient, jump to section 1.3 and return to theory later. - **Deep-read the livestream project (Chapter 3) — it's the master template**: This is the most important chapter because it establishes the complete development workflow (requirements → template design → prompt writing → sandbox testing → deployment) that all five subsequent projects follow. Master the SKILL.md structure here and the rest becomes variation on a theme. - **Use the video QR codes for platform operations**: The book explicitly mentions that basic operations and projects are accompanied by demonstration videos. When you hit UI confusion (e.g., where exactly to upload reference files, how to navigate the sandbox), scan the QR code rather than re-reading text descriptions. - **Don't worry about understanding generated code**: The book repeatedly emphasizes that AI Programming generates scripts (like the stock-scraping crawler) that you don't need to comprehend. Treat code output like you'd treat Excel formulas from an assistant — check the result, not the implementation. If output is wrong, give feedback in natural language. - **Read the SKILL.md examples carefully, not just the narrative**: The actual generated SKILL.md content (task objectives, operation steps, usage examples) shows you the exact format and level of detail expected. When writing your own Skills, mirror this structure — it's the difference between a vague prompt and a reliable expert package. # 【Coverage Limits】 This guide covers the book's opening through the first full project (