Claude Code橙皮书从入门到精通(微信读书特别版) (花叔)（Z-Library）

【One-Line Pitch】 A hands-on, seven-day field manual for turning Claude Code from a terminal tool into your personal engineering team—written for both experienced engineers and total non-coders who want to ship real products by describing what they want, not writing code. 【Book Arc】 - **Opening (~0%–13%)**: Why Claude Code differs from Copilot and Cursor—it's not an autocomplete or IDE extension but a full agent that plans, codes, tests, and commits. Covers installation, account setup, proxy troubleshooting for restricted networks, and the first "Hello, Claude Code!" page. - **Early (~13%–23%)**: The five-step dialogue-driven workflow—describe, review the plan, watch it work, verify output, iterate. Built around a complete CLI project (an AI news aggregator) that exercises network requests, data processing, AI calls, and file output. - **Early (~23%–29%)**: Permission and safety systems: Plan mode for complex tasks, Auto mode with its two-layer AI classifier, `/permissions` whitelists, and Git worktrees for parallel work. Introduces Computer Use (GUI control) and Voice Mode. - **Middle (~29%–42%)**: The CLAUDE.md memory system—the "constitution" for your AI. Covers the guardrails-not-manuals philosophy, the three-level hierarchy (global, project, subdirectory), and Auto Memory that learns from your corrections. - **Late (~42%–100%)**: Advanced topics per the book's structure: Skills extensions, MCP (Model Context Protocol) integration, multi-agent collaboration, and a complete from-requirements-to-launch product case study. *(Excerpts thin out here; details below.)* 【Key Takeaways】 - **The relationship shift matters more than the tech** (Early): Copilot is your input method, Cursor is your pair programmer, but Claude Code is an independent engineer—you become the product manager who says what, not the coder who writes how. This reframing is the book's core thesis. - **Describe outcomes, not code** (Early): The quality of your prompt determines the quality of the output. Give concrete data source URLs, specify the tech stack (TypeScript + tsx), and describe the final artifact (a Markdown report)—then let Claude propose a plan before writing anything. - **The five-step loop is the universal pattern** (Early): Describe → Review plan → Watch execution → Verify output → Iterate. Whether it's a tiny tool or a full product, this cycle repeats. The author's key trick: "Don't write code yet, give me a plan first." - **Permission fatigue is a real security hole** (Early): 93% of permission requests get blindly approved. Auto mode's two-layer classifier (fast pass with 8.5% false positives, deep reasoning bringing it to 0.4%) is the balanced answer—but for production databases or cloud infrastructure, manual confirmation is still wiser. - **CLAUDE.md is a constitution, not a manual** (Middle): Keep it short (~100 lines, ~2500 tokens). Start empty and add one rule each time Claude makes a mistake—this keeps it precise and lean. Rule of thumb: don't write what Claude can read from code; write what it can't guess. - **Avoid the anti-patterns in memory files** (Middle): Don't `@`-reference long documents (they eat context every session); don't write bare "never do X" without offering alternatives; if a command needs paragraphs of explanation, simplify the command itself with a wrapper. - **Voice + Computer Use = keyboard-free building** (Early): Voice Mode for quick task launches and brainstorming, Computer Use for GUI operations like testing a web app or adjusting settings. The combination points toward a future where you describe a product aloud and Claude builds it on screen. - **Parallel work is the agentic superpower** (Early): Boris runs 5 Claude Code instances in separate Git worktrees plus 5–10 web sessions—managing a team of agents rather than doing tasks yourself. This is the workflow shift the book wants you to internalize. 【Reading Tips】 - **Skim the first two chapters** if you're already comfortable with terminals and AI tools—the installation and first-project sections are straightforward. But don't skip the "why" framing in §01; it sets up the mental model for everything after. - **Deep-read the CLAUDE.md chapter** (around 40%): this is where the book earns its keep. The example file and the "guardrails, not manuals" philosophy will save you thousands of tokens and endless frustration. - **Do the news aggregator project, not just read it**: the book is explicitly "open your terminal and follow along." The five-step loop only sticks if you've watched Claude plan, execute, and self-correct at least once. - **Pay attention to the permission section before you get annoyed**: the author predicts you'll hit "approval fatigue" around the 50th `y` press. Knowing Auto mode and `/permissions` exist beforehand will save you from developing bad habits. - **Treat the later chapters as reference, not narrative**: the excerpts suggest the book's final sections (Skills, MCP, multi-agent, full product launch) are practical recipes. Skim them when you need a specific capability rather than reading linearly. 【Coverage Limits】 This guide is based on excerpts covering roughly the first 42% of the book (installation through CLAUDE.md and Auto Memory). The later chapters on Skills extensions, MCP integration, multi-agent collaboration, and the complete product case study are mentioned in the book's structure but not detailed in the available material.