Claude Code The Definitive Guide to Agentic Development (Written by Claude Code etc.)（Z-Library）

# Claude Code: The Definitive Guide to Agentic Development ## 【One-Line Pitch】 A practical field manual for developers who've outgrown Claude Code basics and want to master agentic development—covering mental models, context engineering, multi-agent orchestration, and team adoption strategies. Written entirely by Claude Code itself, this book is for experienced users ready to move from casual prompting to systematic, production-grade workflows. ## 【Book Arc】 - **Opening (~0%–9%)**: Establishes the core mental model shift—Claude Code is an agentic loop, not a chatbot. Explains the 350-line architecture behind the harness, why the context window is the only constraint that matters, and why micromanaging kills productivity. - **Early (~9%–25%)**: Introduces the four-phase workflow (explore, plan, implement, commit) and permission modes as workflow selectors, not just safety dials. Covers the "first-stop" habit—making Claude Code the starting point for every task—and the checkpoint-and-rollback strategy built on clean git states. - **Early (~25%–34%)**: Tackles the collaboration paradox: developers use AI tools in ~60% of work but can fully delegate only 0–20% of tasks. Introduces the thought-partner model and a decision matrix for classifying what to build now, what to hybridize, and what to wait for. - **Middle (~38%–47%)**: Dives into task classification (async vs. sync), session strategy as context management, and the iterative partner model. Emphasizes that a wrong first attempt is data, not failure, and that CLAUDE.md is the persistence layer for cross-session learning. - **Middle (~47%–53%)**: Covers operational practices—commit frequently, revert without hesitation—and the discipline of starting from clean git states. The book's later chapters (beyond the excerpted range) move into advanced capabilities like multi-agent orchestration, MCP integrations, CI/CD automation, and IDE patterns, followed by practical sections on prompt craft, legacy codebases, failure recovery, team adoption, and AI economics. ## 【Key Takeaways】 - **Agentic execution is a loop, not a request-response pair** (Opening): A single prompt can trigger 30–40 iterations of read-plan-act-verify cycles. Internalizing this changes how you interact—you stop asking questions and start giving assignments. - **The context window is the only constraint that matters** (Early): Treat context like a scarce resource. Reading large files, dumping verbose test output, and unnecessary searches all burn expensive context. Proactive compaction and concise CLAUDE.md files (under 500 lines) are survival skills. - **The four-phase workflow—explore, plan, implement, commit—front-loads cheap work** (Early): Exploration and planning in read-only plan mode cost nothing; fixing misunderstandings there beats reverting broken implementations. This is engineering efficiency, not caution. - **Permission modes are workflow selectors, not just safety dials** (Early): They shape how work happens, enabling autonomous runs with periodic check-ins. The directive "commit your work as you go" creates reviewable trails for cherry-picking or reverting. - **Full delegation is a myth—embrace the thought-partner model** (Early): Research shows only 0–20% of tasks can be fully handed off. The practitioners who thrive treat Claude as an iterative collaborator: propose, react, adjust, approve. Give it a rough proposal rather than a blank slate. - **Classify tasks before you start** (Middle): Async candidates (test generation, docs, scaffolding, migrations) have unambiguous success criteria and low risk. Sync candidates (core logic, security, architecture, performance) need supervision. Knowing which is which is itself a skill. - **Session management is context management** (Middle): Sessions aren't persistent environments—a bloated session degrades performance. Fresh sessions with a strong CLAUDE.md often beat continuing stale ones. Fork to explore alternatives without losing progress. - **A wrong first attempt is data, not failure** (Middle): The one-third one-shot success rate is a feature, not a quality metric. Each iteration sharpens Claude's understanding, and multi-session learning accumulates in CLAUDE.md for dramatically better first-pass quality over time. ## 【Reading Tips】 - **Deep-read the Opening and Early sections (0%–34%)**: The mental models here—agentic loop, context economics, four-phase workflow, task classification—are the foundation everything else builds on. Skim if you already practice these, but don't skip the collaboration paradox discussion. - **Treat the Middle section (38%–53%) as your operational playbook**: Session strategy, commit discipline, and the iterative partner model are immediately actionable. Take notes on the async/sync classification matrix and apply it to your current projects. - **Watch for the context-budget mindset**: The book repeatedly emphasizes treating context like compute or memory. If you find yourself skimming, focus on the concrete practices—proactive compaction, concise CLAUDE.md, targeted globs over searches. - **The excerpts don't cover the book's second half in detail**: Multi-agent orchestration, MCP integrations, CI/CD automation, IDE patterns, prompt craft for agentic tools, legacy codebase strategies, failure recovery, team adoption, and AI economics are mentioned but not excerpted. If these matter to you, the full book is worth reading. - **Expect rapid evolution**: The book itself notes details may change as the tool evolves. Treat it as a mindset and pattern guide, and verify specific features against current official documentation. ## 【Coverage Limits】 This guide synthesizes excerpts from roughly the first half of the book (0%–53%). The second half—covering advanced capabilities (multi-agent orchestration, MCP, CI/CD, IDE integration) and practical patterns (prompt craft, legacy codebases, failure recovery, team adoption, economics)—is referenced but not detailed here. ##