Claude Code 的创造者 Boris Cherney 分享了他如何从在 eBay 上卖宝可梦卡到构建增长最快的开发者工具之一，每天提交 20-30 个拉取请求且无需手写一行代码。
Boris Cherney, creator of Claude Code, shares how he went from selling Pokémon cards on eBay to building one of the fastest-growing developer tools, shipping 20-30 pull requests daily with zero handwritten code.
要点 · TL;DR
Claude Code 现在编写了 Anthropic 80% 的代码，工程师无需手动编辑一行。 Claude Code now writes 80% of code at Anthropic, with engineers editing zero lines manually.
印刷机类比：工程师将变成多学科构建者，而不仅仅是编码者。 The printing press analogy: engineers will become multi-disciplinary builders, not just coders.
安全是首要任务；将 Claude Code 发布到实际环境中有助于改进模型安全性。 Safety is the top priority; releasing Claude Code in the wild helps improve model safety.
核心观点 · Key points
像 Claude Code 这样的 AI 编程智能体现在编写了 Anthropic 80% 的代码，工程师无需手动编辑一行。 AI coding agents like Claude Code now write 80% of code at Anthropic, with engineers editing zero lines manually.
印刷机类比：抄写员变成了作者，市场扩大了；工程师将变成多学科构建者。 The printing press analogy: scribes became authors as the market expanded; engineers will become multi-disciplinary builders.
安全是最重要的优先级；在真实环境中发布 Claude Code 有助于研究和改进模型安全性。 Safety is the most important priority; releasing Claude Code in the wild helps study and improve model safety.
原型设计是关键：快速构建多个版本，测试它们，并根据用户反馈迭代。 Prototyping is key: build many versions quickly, test them, and iterate based on user feedback.
代码审查现在使用 AI 作为第一轮，捕获约 80% 的 bug，人类仍然批准最终更改。 Code review now uses AI as first pass, catching ~80% of bugs, with humans still approving final changes.
通才技能——好奇心、适应性、跨学科思维——比单一语言的深度专业知识更有价值。 Generalist skills—curiosity, adaptability, cross-disciplinary thinking—are more valuable than deep expertise in one language.
反共识 · Contrarian takes
Boris 在 Anthropic 的第一个拉取请求被拒绝，因为他手写代码，没有使用 Claude Code。 Boris's first pull request at Anthropic was rejected because he wrote it by hand, not using Claude Code.
Claude Code 最初只是一个简单的 bash 聊天机器人，并非计划中的产品；它从内部使用中自然成长。 Claude Code started as a simple bash chatbot, not a planned product; it grew organically from internal use.
Anthropic 没有头衔——每个人都是“技术员工”——鼓励跨职能协作。 Anthropic has no titles—everyone is 'Member of Technical Staff'—encouraging cross-functional collaboration.
产品需求文档很少被编写；相反，团队发布原型并根据反馈迭代。 PRDs (product requirement documents) are rarely written; instead, teams ship prototypes and iterate from feedback.
Claude Code 的权限系统使用“瑞士奶酪”模型，具有多层安全措施，而非单一解决方案。 Claude Code's permission system uses a 'Swiss cheese' model with multiple safety layers, not a single solution.
Boris 每天编写 20-30 个拉取请求，无需手写代码，模型引入的 bug 比他手动编写时更少。 Boris writes 20-30 pull requests daily without handwritten code, and the model introduces fewer bugs than he would.
本期章节 · Chapters（共 43）
引言与背景Introduction and Background
从健康初创公司学习用户行为Learning from user behavior at a health startup
Sonar 广告集成Sonar ad integration
早期自由职业与第一份工作Early freelancing and first job
Facebook 七年与职业成长Seven years at Facebook and career growth
早期职业生涯与 Facebook 群组Early Career and Facebook Groups
转战 Instagram 与日本Move to Instagram and Japan
Meta 的代码质量与卓越工程Code Quality and Better Engineering at Meta
衡量工程师生产力Measuring Engineer Productivity
加入 Anthropic 与首个 PR 被拒Joining Anthropic and First PR Rejected
赞助商插播：StatsigSponsor Break: Statsig
Claude Code 的起源Origin of Claude Code
从聊天机器人到智能体 AIFrom Chatbot to Agentic AI
发布审查与采用Launch Review and Adoption
转向信任 Claude CodeSwitch to Trusting Claude Code
开发工作流与技巧Development Workflow and Tips
iOS 应用与并行智能体iOS app and parallel agents
用 Lint 规则自动化繁琐工作Automating Tedious Work with Lint Rules
何时跳过人工审查When to Skip Human Review
确定性检查与非确定性检查Deterministic vs Non-Deterministic Checks
赞助商插播：WorkOSSponsor Break: WorkOS
Claude Code 架构与安全层Claude Code architecture and safety layers
RAG 与 Agent-X 搜索对比RAG vs Agent-X search
权限系统与沙箱Permission system and sandboxing
权限系统与安全Permission System and Safety
头衔与通才文化Titles and Generalist Culture
PRD 与产品流程PRDs and Product Process
Anthropic 的原型文化Prototyping culture at Anthropic
Claude Colab：可视化智能体界面Claude Colab: A Visual Agent Interface
Claude Colab 背后的复杂性Complexity Behind Claude Colab
技术栈与平台选择Tech stack and platform choice
可观测性与隐私Observability and privacy
从 Cowerk 早期学到的经验Early learnings from Cowerk
智能体团队与非相关上下文窗口Agent teams and uncorrelated context windows
Opus 4.6 与群体智能Opus 4.6 and Swarm Intelligence
拥抱 AI 的快速变化Embracing Rapid Change in AI
放下编码身份Letting Go of Coding Identity
编码作为手艺的失落与哀伤Grief and Loss of Coding as a Craft
印刷术类比The Printing Press Analogy
印刷术类比与 AI 影响Printing Press Analogy for AI Impact
对 AI 安全信念的改变Changed Beliefs on AI Safety
重要技能与被抛弃的技能Skills That Matter and Those Left Behind