{"uuid": "23cfce2e-7c2d-4e72-a19d-0ed7ea3e57c2", "vulnerability_lookup_origin": "1a89b78e-f703-45f3-bb86-59eb712668bd", "author": "9f56dd64-161d-43a6-b9c3-555944290a09", "vulnerability": "CVE-2025-54795", "type": "seen", "source": "https://gist.github.com/yurukusa/ebe57afa9cdd9363bb0cba15f5c51d7e", "content": "# Claude Code reliability books in May 2026: a market overview\n\nFour books targeting Claude Code operators, engineers, and reliability concerns shipped or are shipping in the first half of May 2026. They differ structurally \u2014 not just in length, price, or language, but in what they actually claim to be useful for. If you're trying to figure out which one(s) to buy, the question worth asking is not \"which is best\" but \"which solves the problem I actually have.\" This overview is for that question.\n\nI'm the author of one of the four. I've tried to write the rest of this piece in a way that holds up if you read it before learning that, and to make the comparison useful even for people whose answer turns out to be a different book.\n\n---\n\n## The four books, in shipping order\n\n### 1. Greg Lim \u2014 Claude Code Crash Course: Build Real-World Apps with AI\n\nShipped April 9, 2026. 186 pages. Amazon Kindle and paperback. Four ratings, all five-star, no detailed reviews yet.\n\nThe \"Crash Course\" series is Greg Lim's signature format \u2014 he has parallel books for Git/GitHub, Ollama, Claude 3, and several other tools, each oriented around a developer who wants to ship a working application without spending three weeks reading the official documentation. The audience is someone who has heard of Claude Code, wants to use it for a real project, and needs a guided path from install to first deployed application. Failure cases and incident analysis are not in scope.\n\nIf your situation is \"I have not yet used Claude Code on a real project and I want a structured 186-page on-ramp,\" this is a reasonable choice. Amazon distribution makes it available everywhere; the Crash Course series's consistent format is a known quantity if you've used Lim's other books.\n\n### 2. \u30a4\u30f3\u30d7\u30ec\u30b9 \u2014 \u5b8c\u5168\u89e3\u8aac! Claude Code \u30c6\u30af\u30cb\u30c3\u30af\u96c6 \u7206\u901f\u958b\u767a\u306e\u305f\u3081\u306e\u5b8c\u5168\u30ac\u30a4\u30c9\n\nShipped May 8, 2026. Japanese-language print and ebook. Print 4,000 JPY, ebook 3,800 JPY.\n\nComprehensive practical guide oriented around \"vibe coding\" \u2014 the workflow of building applications by describing what you want and iterating with the model. Published by \u30a4\u30f3\u30d7\u30ec\u30b9, an established Japanese technical publisher whose name carries credibility in the Japanese developer market.\n\nIf you read Japanese and want a comprehensive guide-style book with print availability and a recognized publisher's editorial standard, this is the strongest option. The focus is on \"how to use Claude Code productively,\" not on failure modes or reliability engineering.\n\n### 3. Thomas De Vos \u2014 Claude Code: Building Production Agents That Actually Work\n\nCurrently 93% complete on Leanpub, last updated May 11, 2026. 493 pages, 31 chapters. Minimum $9.99, suggested $29.00.\n\nThe author has built AI systems for regulated financial institutions for over a decade and writes for \"Senior AI engineers, technical leads, and architects evaluating Claude Code for production use.\" The chapter list is the strongest signal of fit: agent loop, tools, hooks, MCP, the Agent SDK, permissions, sandboxing, network egress and secrets, policy as code, evals, observability, failure modes and reliability engineering, cost engineering, team workflows. Chapter 26 is specifically \"Failure modes and reliability engineering\" with multiple worked examples including sycophanticity in screening agents and a \"sanctions plugin phoning home\" incident.\n\nThe orientation is engineering \u2014 what to build, how to build it correctly, how to operate it under SLOs and error budgets. Worked examples come from regulated financial environments. The 493-page length is unusual for a Claude Code book and reflects the systematic coverage.\n\nIf your situation is \"I am building or evaluating Claude Code for production deployment in a regulated environment and I need a systematic engineering reference,\" this is the book. The engineering depth and the regulated-financial perspective are differentiating features no other book in the market currently offers.\n\n### 4. Yurukusa \u2014 Claude Code Claim-Verify Handbook\n\nShips May 22, 2026, on Gumroad. 89 pages, $19. Free preview Gist available before launch.\n\nA forensic catalog of 130 cases (15 in the body, 115 in Appendix D) drawn directly from the `anthropics/claude-code` issue tracker, where the assistant or tool emitted a \"verified\" or \"completed\" or \"set\" status surface while the underlying runtime did something else. The orientation is operator-side \u2014 you already run Claude Code, you're hitting unexplained failures, you need to figure out which of the documented failure modes you're inside.\n\nThe book provides: a three-stage diagnostic framework (operator intent \u2192 status claim \u2192 runtime action) for triaging your own session, fourteen user-side defenses (hooks, audit scripts, configuration patterns), five automated detection tools with implementation and 165+ test cases passing, and a continuing-evidence log that documents the cluster's acceleration from a baseline of 0.37 cases/day in April to roughly 13 cases/day across the May 9\u201318 window.\n\nIndustry validation in the book includes Anthropic's own admission of approval fatigue (March 25 engineering blog), three CVEs (CVE-2026-33068, CVE-2025-54795, CVE-2026-39861), the leaked v2.1.88 source code with the internal benchmark showing 29-30% false claims rate for Capybara v8 (regressed from 16.7% at v4), and Anthropic's own published C compiler experiment in which sixteen parallel Claude agents over two weeks and $20,000 produced code slower than GCC at `-O0` with \"new features and bugfixes frequently broke existing functionality\" stated by the engineering team itself.\n\nIf your situation is \"I am already running Claude Code, I am losing money and trust to silent failures, and I need to figure out which of the known failure modes my session is hitting right now,\" this is the book.\n\n---\n\n## Three structural approaches, not four books\n\nGreg Lim, \u30a4\u30f3\u30d7\u30ec\u30b9, and the comprehensive guide tradition share an approach: explain what the tool is and how to use it, oriented around a reader who hasn't fully adopted it yet. The differences between Lim's Crash Course and \u30a4\u30f3\u30d7\u30ec\u30b9's complete guide are real but situated within a shared \"introduction\" frame.\n\nDe Vos's book is a different approach. It assumes adoption is decided and asks \"how do you build and operate this in production?\" The 31-chapter structure is the engineering equivalent of a systems administration manual \u2014 you don't read it cover to cover; you go to the chapter for the problem you're hitting.\n\nThe Handbook is a third approach. It assumes both adoption and operation are decided and asks \"when something goes wrong silently and the tool reports success, how do you triage what actually happened?\" The structure is forensic \u2014 130 cases organized by failure mode, with reproduction steps and detection paths, plus the framework for applying the same analysis to cases the book doesn't cover.\n\nA reader needing only the first approach should buy Lim or \u30a4\u30f3\u30d7\u30ec\u30b9. A reader needing the second should buy De Vos. A reader needing the third should buy the Handbook. Readers operating in production at scale will likely want both De Vos and the Handbook \u2014 they are complementary, not substitutable. De Vos describes the engineering posture; the Handbook describes the 130 documented operator-side failures that any production deployment will eventually encounter.\n\n---\n\n## Where the books overlap and where they don't\n\nThe chapter titles give a clean picture of where overlap exists.\n\nDe Vos's Chapter 26, \"Failure modes and reliability engineering,\" and the Handbook's body cover the same general territory \u2014 what goes wrong with Claude Code in production and how to detect or prevent it. The treatments differ structurally. De Vos's chapter is a survey within a 31-chapter engineering manual; the Handbook is 89 pages dedicated to documenting and triaging this one category of failure. If you're choosing between them on this single chapter alone, the question is how much depth you need on this specific category versus how much you need the surrounding 30 chapters of engineering context.\n\nDe Vos's chapters on the SDK, MCP, permissions, sandboxing, and policy-as-code have no counterpart in the Handbook \u2014 those are engineering topics, not operator-side forensic categories. If you're building rather than operating, those chapters are the value, and the Handbook is not what you need.\n\nThe Handbook's continuing-evidence log (the 115 cases in Appendix D, observed across 233 hours from May 9\u201318) has no counterpart in De Vos. The empirical density is the differentiator \u2014 every case is a specific GitHub issue with the operator's reported behavior, the runtime's actual behavior, and the divergence framework applied. If you want to see what the failure pattern looks like across 130 actual reports rather than the worked examples in De Vos's chapter, the Handbook is where that lives.\n\n---\n\n## A note on independent verification\n\nOne thing worth flagging across all four books: only one of them \u2014 the Handbook \u2014 claims to be the operator-side organized record of a problem that Anthropic, three CVE authorities, four security publications, and Anthropic's own engineering team's published self-experiment all independently acknowledge. This is not a marketing claim; it's a structural statement about the cluster the Handbook catalogs. The other three books either don't address this specific category (Lim, \u30a4\u30f3\u30d7\u30ec\u30b9) or treat it as one chapter within a broader engineering framework (De Vos).\n\nFor readers who care about whether the book is documenting a real pattern versus describing an isolated set of incidents: the Handbook's eight independent verification axes (security publications, CVEs, Anthropic's own blog, the 860-point HN production-database-deletion thread, the v2.1.88 source code leak, the Brodzinski outside-editor piece, the Anthropic C compiler self-experiment, and the Zerostack alternative-tool emergence) are the verification structure. The other three books don't engage this question because their structure doesn't require it \u2014 a \"how to use Claude Code\" book or a \"how to build production agents\" book operates correctly without taking a position on whether the underlying tool's reliability claims match operator-side reality.\n\n---\n\n## Recommendations by situation\n\nIf you have not yet used Claude Code on a real project: Greg Lim's Crash Course (English) or \u30a4\u30f3\u30d7\u30ec\u30b9's \u30c6\u30af\u30cb\u30c3\u30af\u96c6 (Japanese). Both will get you productive in a structured way.\n\nIf you are building Claude Code agents for production deployment and need engineering depth: Thomas De Vos's Building Production Agents. The 31-chapter breadth and the regulated-financial perspective are the differentiating value.\n\nIf you are already operating Claude Code and hitting unexplained silent failures: the Claim-Verify Handbook. The 130-case forensic catalog and the three-stage triage framework are the differentiating value.\n\nIf you are operating at scale in production: De Vos plus the Handbook. They are complementary \u2014 engineering posture from De Vos, operator-side failure triage from the Handbook.\n\nIf you read Japanese and want a print-format comprehensive guide: \u30a4\u30f3\u30d7\u30ec\u30b9's \u30c6\u30af\u30cb\u30c3\u30af\u96c6.\n\n---\n\n## Pricing context\n\nLim: $9.99\u201314.99 Kindle range typical for the Crash Course series, paperback varies.\n\u30a4\u30f3\u30d7\u30ec\u30b9: 3,800 JPY ebook, 4,000 JPY print (approximately $25\u201327 USD).\nDe Vos: $9.99 minimum, $29.00 suggested on Leanpub.\nHandbook: $19 on Gumroad, no minimum, free preview Gist available.\n\nIf price is the binding constraint, the Italian Leanpub guide (Claude Code: Guida pratica, $0 minimum / $6 suggested, 140 pages, Italian) is worth mentioning as a fifth option not detailed above \u2014 Creative Commons licensing and pay-what-you-want make it the lowest-friction entry point for anyone who reads Italian. It does not cover the same material as the four books above; it's a comprehensive practical guide oriented at first-time users.\n\n---\n\n## What this overview does not do\n\nThis overview does not rank the books. Lim's book is the right purchase for someone who needs Lim's book; De Vos's is the right purchase for someone who needs De Vos's; the Handbook is the right purchase for someone who needs the Handbook. \"Best Claude Code book of 2026\" is a category that does not exist because the books are not in the same category. The question worth asking is which problem you have, and the four-way split above is the structural shape of the market answering that question.\n\nThe market itself is a useful signal: four books published or shipping within a six-week window, three of them in English, one in Japanese, three approaches (introduction, engineering, forensic). That this is the shape of the market in May 2026 \u2014 rather than, say, two introduction-style books competing on quality \u2014 is itself information about where Claude Code is in its adoption curve. The tool is past the \"what is this\" phase and into the \"how do I deploy and operate this without losing money and trust\" phase. The book market reflects that shift.\n\n---\n\n## Disclosure and self-reference\n\nI wrote the Claim-Verify Handbook. The Gist with the free preview is at https://gist.github.com/yurukusa/5242a540c43769df76a448269e2f182b and the launch page is at https://yurukusa.gumroad.com/l/claim-verify-handbook (ships May 22, 2026, $19). The rest of this overview is structured to be useful regardless of which book you end up buying, and the recommendations above represent my honest read on which book solves which problem. If your situation maps to Lim, \u30a4\u30f3\u30d7\u30ec\u30b9, or De Vos, those are the right purchases for that situation, and the Handbook is not what you need.\n\nThe market analysis above draws from each book's published sales page and the publisher's distribution information as of May 18, 2026. Page counts, prices, and release statuses are accurate as of that date.\n", "creation_timestamp": "2026-05-18T06:38:21.000000Z"}