这是用户在 2025-6-25 18:21 为 https://app.immersivetranslate.com/pdf-pro/d0c5bc08-3073-4940-a480-2af8dc256c4d/ 保存的双语快照页面，由 沉浸式翻译 提供双语支持。了解如何保存？

ANTHROPIC

How Anthropic teams use Claude Code
Anthropic 团队如何使用 Claude Code

Anthropic’s internal teams are transforming their workflows with Claude Code, enabling developers and non-technical staff to tackle complex projects, automate tasks, and bridge skill gaps that previously limited their productivity.
Anthropic 的内部团队正在通过 Claude Code 改造他们的工作流程，使开发人员和非技术人员能够处理复杂项目、自动化任务，并弥合此前限制他们生产力的技能差距。
Through interviews with our own Claude Code power users, we’ve gathered insights on how different departments leverage Claude Code, its impact on their work, and tips for other organizations considering adoption.
通过对我们自己的 Claude Code 高级用户的访谈，我们收集了不同部门如何利用 Claude Code、其对工作影响的见解，以及对其他考虑采用该工具的组织的建议。

Contents  目录

Claude Code for data infrastructure … 3
用于数据基础设施的 Claude Code … 3
Claude Code for product development … 5
用于产品开发的 Claude Code … 5
Claude Code for security engineering … 7
用于安全工程的 Claude Code … 7
Claude Code for inference … 9
Claude Code 用于推理 … 9
Claude Code for data science and visualization … 11
Claude Code 用于数据科学与可视化 … 11
Claude Code for API … 13
Claude Code 用于 API … 13
Claude Code for growth marketing … 15
Claude Code 用于增长营销 … 15
Claude Code for product design … 17
Claude Code 用于产品设计 … 17
Claude Code for RL engineering … 19
Claude Code 用于强化学习工程 … 19
Claude Code for legal … 21
Claude Code 用于法律 … 21

Claude Code for data infrastructure
Claude Code 用于数据基础设施

The Data Infrastructure team organizes all business data for teams across the company.
数据基础设施团队负责为公司各团队组织所有业务数据。
They use Claude Code for automating routine data engineering tasks, troubleshooting complex infrastructure issues, and creating documented workflows for technical and non-technical team members to access and manipulate data independently.
他们使用 Claude Code 自动化常规数据工程任务，排查复杂的基础设施问题，并创建有文档的工作流程，供技术和非技术团队成员独立访问和操作数据。

Main Claude Code use cases
主要的 Claude Code 使用案例

Kubernetes debugging with screenshots
带截图的 Kubernetes 调试

When Kubernetes clusters went down and weren’t scheduling new pods, the team used Claude Code to diagnose the issue.
当 Kubernetes 集群宕机且无法调度新 pod 时，团队使用 Claude Code 诊断问题。
They fed screenshots of dashboards into Claude Code, which guided them through Google Cloud’s UI menu by menu until they found a warning indicating pod IP address exhaustion.
他们将仪表盘的截图输入 Claude Code，Claude Code 逐步引导他们通过 Google Cloud 的 UI 菜单，直到发现一个指示 pod IP 地址耗尽的警告。
Claude Code then provided the exact commands to create a new IP pool and add it to the cluster, bypassing the need to involve networking specialists.
随后，Claude Code 提供了创建新 IP 池并将其添加到集群的准确命令，避免了需要网络专家介入。

Plain text workflows for finance team
财务团队的纯文本工作流程

The team showed finance team members how to write plain text files describing their data workflows, then load them into Claude Code to get fully automated execution.
团队向财务团队成员展示了如何编写描述其数据工作流程的纯文本文件，然后将其加载到 Claude Code 中以实现完全自动化执行。
Employees with no coding experience could describe steps like “query this dashboard, get information, run these queries, produce Excel output,” and Claude Code would execute the entire workflow, including asking for required inputs like dates.
没有编码经验的员工可以描述诸如“查询此仪表板，获取信息，运行这些查询，生成 Excel 输出”之类的步骤，Claude Code 会执行整个工作流程，包括请求所需的输入，如日期。

Codebase navigation for new hires
新员工的代码库导航

When new data scientists join the team, they’re directed to use Claude Code to navigate their massive codebase. Claude Code reads their Claude.md files (documentation), identifies relevant files for specific tasks, explains data pipeline dependencies, and helps newcomers understand which upstream sources feed into dashboards. This replaces traditional data catalogs and discoverability tools.
当新的数据科学家加入团队时，他们会被引导使用 Claude Code 来导航庞大的代码库。Claude Code 会读取他们的 Claude.md 文件（文档），识别特定任务的相关文件，解释数据管道依赖关系，并帮助新成员了解哪些上游来源为仪表板提供数据。这取代了传统的数据目录和可发现性工具。

End-of-session documentation updates
会话结束时的文档更新

The team asks Claude Code to summarize completed work sessions and suggest improvements at the end of each task. This creates a continuous improvement loop where Claude Code helps refine the Claude.md documentation and workflow instructions based on actual usage, making subsequent iterations more effective.
团队在每个任务结束时请求 Claude Code 总结已完成的工作会话并提出改进建议。这创建了一个持续改进的循环，Claude Code 根据实际使用情况帮助完善 Claude.md 文档和工作流程说明，使后续迭代更加高效。

Parallel task management across multiple instances
跨多个实例的并行任务管理

When working on long-running data tasks, they open multiple instances of Claude Code in different repositories for different projects.
在处理长时间运行的数据任务时，他们会在不同项目的不同代码库中打开多个 Claude Code 实例。
Each instance maintains full context, so when they switch back after hours or days, Claude Code remembers exactly what they were doing and where they left off, enabling true parallel workflow management without context loss.
每个实例都保持完整的上下文，因此当他们在数小时或数天后切换回来时，Claude Code 能准确记住他们之前的操作和停留的位置，实现真正的并行工作流管理而不会丢失上下文。

Claude Code for data infrastructure
用于数据基础设施的 Claude Code

Team impact  团队影响

Resolved infrastructure problems without specialized expertise
无需专业知识即可解决基础设施问题

Resolved Kubernetes cluster issues that would normally require pulling in systems or networking team members, using Claude Code to diagnose problems and provide exact fixes.
解决了通常需要调动系统或网络团队成员的 Kubernetes 集群问题，使用 Claude Code 诊断问题并提供精确的修复方案。

Accelerated onboarding  加快了入职流程

New data analysts and team members can quickly understand complex systems and contribute meaningfully without extensive guidance.
新的数据分析师和团队成员能够快速理解复杂系统，并在无需大量指导的情况下做出有意义的贡献。

Enhanced support workflow
优化了支持工作流程

Can process much larger data volumes and identify anomalies (like monitoring 200 dashboards) that would be impossible for humans to review manually.
能够处理更大规模的数据量并识别异常（例如监控 200 个仪表盘），这是人工手动审查不可能完成的。

Enabled cross-team self-service
实现了跨团队自助服务

Finance teams with no coding experience can now execute complex data workflows independently.
没有编码经验的财务团队现在可以独立执行复杂的数据工作流。

Top tips from the Data Infrastructure team
数据基础设施团队的顶级建议

Write detailed Claude.md files
编写详细的 Claude.md 文件

The better you document your workflows, tools, and expectations in Claude.md files, the better Claude Code performs. This made Claude Code excel at routine tasks like setting up new data pipelines when you have existing patterns.
你在 Claude.md 文件中对工作流程、工具和期望的记录越详细，Claude Code 的表现就越好。这使得 Claude Code 在处理已有模式的新数据管道设置等常规任务时表现出色。

Use MCP servers instead of CLI for sensitive data
对于敏感数据，使用 MCP 服务器而非 CLI

They recommend using MCP servers rather than the BigQuery CLI to maintain better security control over what Claude Code can access, especially for handling sensitive data that requires logging or has potential privacy concerns.
他们建议使用 MCP 服务器而不是 BigQuery CLI，以更好地控制 Claude Code 的访问权限，特别是在处理需要日志记录或存在潜在隐私问题的敏感数据时。

Share team usage sessions
分享团队使用会话

The team held sessions where members demonstrated their Claude Code workflows to each other. This helped spread best practices and showed different ways to use the tool they might not have discovered on their own.
团队举办了成员相互展示 Claude Code 工作流程的会话。这有助于传播最佳实践，并展示他们可能自己未曾发现的使用该工具的不同方法。

Claude Code for product development
用于产品开发的 Claude Code

The Claude Code team uses their own product to build updates to Claude Code, expanding the product’s enterprise capabilities and agentic loop functionalities.
Claude Code 团队使用他们自己的产品来构建 Claude Code 的更新，扩展产品的企业功能和自主循环功能。

Main Claude Code use cases
Claude Code 主要使用场景

Fast prototyping with auto-accept mode
启用自动接受模式进行快速原型设计

Engineers use Claude Code for rapid prototyping by enabling “autoaccept mode” (shift+tab) and setting up autonomous loops where Claude writes code, runs tests, and iterates continuously.
工程师通过启用“自动接受模式”（shift+tab）并设置自主循环，让 Claude 编写代码、运行测试并持续迭代，从而实现快速原型设计。
They give Claude abstract problems they’re unfamiliar with, let it work autonomously, then review the 80 % 80 % 80%80 \% complete solution before taking over for final refinements. Teams emphasize starting from a clean git state and committing checkpoints regularly so they can easily revert any incorrect changes if Claude goes off track.
他们将不熟悉的抽象问题交给 Claude 自主处理，然后在接管进行最终完善之前，审查 Claude 完整的解决方案。团队强调从干净的 git 状态开始，并定期提交检查点，以便在 Claude 偏离轨道时轻松回滚任何错误更改。

Synchronous coding for core features
核心功能的同步编码

For more critical features touching the application’s business logic, the team works synchronously with Claude Code, giving detailed prompts with specific implementation instructions.
对于涉及应用业务逻辑的更关键功能，团队与 Claude Code 同步工作，提供带有具体实现指令的详细提示。
They monitor the process in real-time to ensure code quality, style guide compliance, and proper architecture while letting Claude handle the repetitive coding work.
他们实时监控过程，以确保代码质量、风格指南的遵守以及正确的架构，同时让 Claude 处理重复的编码工作。

Building Vim mode  构建 Vim 模式

One of their most successful async projects was implementing Vim key bindings for Claude Code. They asked Claude to build the entire feature (despite it not being a priority), and roughly 70 % 70 % 70%70 \% of the final implementation came from Claude’s autonomous work, requiring only a few iterations to complete.
他们最成功的异步项目之一是为 Claude Code 实现 Vim 键绑定。他们让 Claude 构建整个功能（尽管这不是优先事项），最终实现的大约 70 % 70 % 70%70 \% 来自 Claude 的自主工作，仅需几次迭代即可完成。

Test generation and bug fixes
测试生成和错误修复

They use Claude Code to write comprehensive tests after implementing features and handle simple bug fixes identified in pull request reviews.
他们使用 Claude Code 在实现功能后编写全面的测试，并处理在拉取请求审查中发现的简单错误修复。
They also leverage GitHub Actions integration to have Claude automatically address Pull Request comments like formatting issues or function renaming.
他们还利用 GitHub Actions 集成，让 Claude 自动处理拉取请求评论中的格式问题或函数重命名等事项。

Codebase exploration  代码库探索

When working with unfamiliar codebases (like the monorepo or API side), the team uses Claude Code to quickly understand how systems work.
在处理不熟悉的代码库（如 monorepo 或 API 端）时，团队使用 Claude Code 快速了解系统的工作原理。
Instead of waiting for Slack responses, they ask Claude directly for explanations and code references, saving significant time in context switching.
他们不再等待 Slack 的回复，而是直接向 Claude 询问解释和代码参考，节省了大量在上下文切换上的时间。

Claude Code for product development
用于产品开发的 Claude Code

Team impact  团队影响

Faster feature implementation
更快的功能实现

Successfully implemented complex features like Vim mode with 70% of code written autonomously by Claude.
成功实现了复杂功能，如 Vim 模式，其中 70% 的代码由 Claude 自动编写。

Improved development velocity
提升开发速度

Can rapidly prototype features and iterate on ideas without getting bogged down in implementation details.
能够快速原型设计功能并迭代想法，而不会陷入实现细节。

Enhanced code quality through automated testing
通过自动化测试提升代码质量

Claude generates comprehensive tests and handles routine bug fixes, maintaining high standards while reducing manual effort.
Claude 生成全面的测试并处理常规的错误修复，在减少人工工作量的同时保持高标准。

Better codebase exploration
更好的代码库探索

Team members can quickly understand unfamiliar parts of the monorepo without waiting for colleague responses.
团队成员可以快速理解单一代码库中不熟悉的部分，无需等待同事回复。

Top tips from the Claude Code team
Claude Code 团队的顶级技巧

Create self-sufficient loops
创建自给自足的循环

Set up Claude to verify its own work by running builds, tests, and lints automatically. This allows Claude to work longer autonomously and catch its own mistakes, especially effective when you ask Claude to generate tests before writing code.
设置 Claude 通过自动运行构建、测试和代码风格检查来验证自己的工作。这使 Claude 能够更长时间自主工作并发现自身错误，尤其在你让 Claude 先生成测试再编写代码时效果显著。

Develop task classification intuition
培养任务分类直觉

Learn to distinguish between tasks that work well asynchronously (peripheral features, prototyping) versus those needing synchronous supervision (core business logic, critical fixes).
学会区分适合异步处理的任务（外围功能、原型设计）与需要同步监督的任务（核心业务逻辑、关键修复）。
Abstract tasks on the product’s edges can be handled with “auto-accept mode,” while core functionality requires closer oversight.
产品边缘的抽象任务可以通过“自动接受模式”处理，而核心功能则需要更严格的监督。

Form clear, detailed prompts
形成清晰、详细的提示

When components have similar names or functions, be extremely specific in your requests. The better and more detailed your prompt, the more you can trust Claude to work independently without unexpected changes to the wrong parts of the codebase.
当组件名称或功能相似时，务必在请求中非常具体。提示越好越详细，就越能信任 Claude 独立工作，而不会对代码库的错误部分进行意外更改。

Claude Code for security engineering
Claude Code 用于安全工程

The Security Engineering team focuses on securing the software development lifecycle, supply chain security, and development environment security. They use Claude Code extensively for writing and debugging code.
安全工程团队专注于保障软件开发生命周期、安全供应链以及开发环境的安全。他们广泛使用 Claude Code 进行代码编写和调试。

Main Claude Code use cases
Claude Code 主要使用场景

Complex infrastructure debugging
复杂基础设施调试

When working on incidents, they feed Claude Code stack traces and documentation, asking it to trace control flow through the codebase.
在处理事件时，他们向 Claude Code 提供堆栈跟踪和文档，要求它追踪代码库中的控制流。
This significantly reduces time-to-resolution for production issues, allowing them to understand problems that would normally take 10-15 minutes of manual code scanning in about 5 minutes.
这大大缩短了生产问题的解决时间，使他们能够在大约 5 分钟内理解通常需要 10-15 分钟手动代码扫描的问题。

Terraform code review and analysis
Terraform 代码审查与分析

For infrastructure changes requiring security approval, they copy Terraform plans into Claude Code to ask “what’s this going to do?
对于需要安全审批的基础设施变更，他们将 Terraform 计划复制到 Claude Code 中，询问“这将做什么？”
Am I going to regret this?” This creates tighter feedback loops and makes it easier for the security team to quickly review and approve infrastructure changes, reducing bottlenecks in the development process.
“我会后悔这样做吗？”这创造了更紧密的反馈循环，使安全团队能够更快地审查和批准基础设施变更，减少开发过程中的瓶颈。

Documentation synthesis and runbooks
文档综合与运行手册

They have Claude Code ingest multiple documentation sources and create markdown runbooks, troubleshooting guides, and overviews.
他们让 Claude Code 吸收多个文档来源，创建 markdown 格式的运行手册、故障排除指南和概述。
They use these condensed documents as context for debugging real issues, creating a more efficient workflow than searching through full knowledge bases.
他们使用这些精简的文档作为调试实际问题的上下文，创建了比搜索完整知识库更高效的工作流程。

Test-driven development workflow
测试驱动开发工作流程

Instead of their previous “design doc → → rarr\rightarrow janky code → → rarr\rightarrow refactor → → rarr\rightarrow give up on tests” pattern, they now ask Claude Code for pseudocode, guide it through test-driven development, and periodically check in to steer it when stuck, resulting in more reliable and testable code.
他们不再采用之前的“设计文档 → → rarr\rightarrow 代码杂乱无章 → → rarr\rightarrow 重构 → → rarr\rightarrow 放弃测试”的模式，而是向 Claude Code 请求伪代码，通过测试驱动开发引导它，并在遇到困难时定期检查以引导方向，从而生成更可靠且易于测试的代码。

Context switching and project onboarding
上下文切换与项目入门

When contributing to existing projects like “dependant” (a web application for security approval workflows), they use Claude Code to write, review, and execute specifications written in markdown and stored in the codebase, enabling meaningful contributions within days instead of weeks.
在为现有项目如“dependant”（一个用于安全审批工作流的网络应用）做贡献时，他们使用 Claude Code 编写、审查并执行以 markdown 编写并存储在代码库中的规范，使得能够在几天内而非数周内做出有意义的贡献。

Claude Code for security engineering
Claude Code 用于安全工程

Team impact  团队影响

Reduced incident resolution time
缩短事件解决时间

Infrastructure debugging that normally takes 10-15 minutes of manual code scanning now takes about 5 minutes.
通常需要 10-15 分钟手动代码扫描的基础设施调试，现在大约只需 5 分钟。

Improved security review cycle
改进的安全审查周期

Terraform code reviews for security approval happen much faster, eliminating developer blocks while waiting for security team approval.
Terraform 代码的安全审批审查速度大大加快，消除了开发人员在等待安全团队批准时的阻塞。

Enhanced cross-functional contribution
增强的跨职能贡献

Team members can meaningfully contribute to projects within days instead of weeks of context building.
团队成员可以在几天内而非数周的背景构建后，对项目做出有意义的贡献。

Better documentation workflow
更好的文档工作流程

Synthesized troubleshooting guides and runbooks from multiple sources create more efficient debugging processes.
从多个来源综合的故障排除指南和运行手册创造了更高效的调试流程。

Top tips from the Security Engineering team
安全工程团队的顶级建议

Use custom slash commands extensively
广泛使用自定义斜杠命令

Security engineering uses 50 % 50 % 50%50 \% of all custom slash command implementations in the entire monorepo. These custom commands streamline specific workflows and speed up repeated tasks.
安全工程使用了整个单一代码库中所有自定义斜杠命令实现的 50 % 50 % 50%50 \% 。这些自定义命令简化了特定的工作流程，加快了重复任务的执行速度。

Let Claude talk first
让 Claude 先说话

Instead of asking targeted questions for code snippets, they now tell Claude Code to “commit your work as you go” and let it work autonomously with periodic check-ins, resulting in more comprehensive solutions.
他们不再针对代码片段提出具体问题，而是告诉 Claude Code“边做边提交你的工作”，让它自主工作并定期检查，从而产生更全面的解决方案。

Leverage it for documentation
利用它来编写文档

Beyond coding, Claude Code excels at synthesizing documentation and creating structured outputs. They provide writing samples and formatting preferences to get documents they can immediately use in Slack, Google Docs, and other tools to avoid interface switching fatigue.
除了编码，Claude Code 在综合文档和创建结构化输出方面表现出色。用户可以提供写作样本和格式偏好，生成可直接在 Slack、Google Docs 及其他工具中使用的文档，避免频繁切换界面带来的疲劳。

Claude Code for inference
Claude Code 用于推理

The Inference team manages the memory system that stores information while Claude reads your prompt and generates its response. Team members, especially those who are new to machine learning, can use Claude Code extensively to bridge that knowledge gap and accelerate their work.
推理团队管理着在 Claude 读取提示并生成响应时存储信息的内存系统。团队成员，尤其是那些刚接触机器学习的新手，可以广泛使用 Claude Code 来弥补知识差距，加快工作进度。

Main Claude Code use cases
Claude Code 的主要使用场景

Codebase comprehension and onboarding
代码库理解与入门

The team relies heavily on Claude Code to quickly understand the architecture when joining a complex codebase.
团队在加入复杂代码库时，严重依赖 Claude Code 快速理解架构。
Instead of manually searching GitHub repos, they ask Claude to find which files call specific functionalities, getting results in seconds rather than asking colleagues or searching manually.
他们不再手动搜索 GitHub 仓库，而是让 Claude 查找调用特定功能的文件，几秒钟内即可获得结果，而无需询问同事或手动搜索。

Unit test generation with edge case coverage
包含边界情况覆盖的单元测试生成

After writing core functionality, they ask Claude to write comprehensive unit tests. Claude automatically includes missed edge cases, completing what would normally take significant mental energy in minutes, acting like a coding assistant they can review.
在编写核心功能后，他们让 Claude 编写全面的单元测试。Claude 会自动包含遗漏的边界情况，完成通常需要大量脑力的工作，仅用几分钟，就像一个可以审查的编码助手。

Machine learning concept explanation
机器学习概念解释

Without a machine learning background, team members depend on Claude to explain model-specific functions and settings. What would require an hour of Google searching and reading documentation now takes 10-20 minutes, reducing research time by 80%.
没有机器学习背景的团队成员依赖 Claude 来解释模型特定的函数和设置。原本需要一小时的 Google 搜索和阅读文档，现在只需 10 到 20 分钟，研究时间减少了 80%。

Cross-language code translation
跨语言代码翻译

When testing functionality in different programming languages, they explain what they want to test and Claude writes the logic in the required language (like Rust), eliminating the need to learn new languages just for testing purposes.
在测试不同编程语言的功能时，他们说明想要测试的内容，Claude 会用所需的语言（如 Rust）编写逻辑，免去了仅为测试目的而学习新语言的需求。

Command recall and Kubernetes management
命令回顾与 Kubernetes 管理

Instead of remembering complex Kubernetes commands, they ask Claude for the correct syntax, like “how to get all pods or deployment status,” and receive the exact commands needed for their infrastructure work.
他们不必记住复杂的 Kubernetes 命令，而是向 Claude 询问正确的语法，比如“如何获取所有 pod 或部署状态”，并获得完成基础设施工作的准确命令。

Claude Code for inference
用于推理的 Claude Code

Team impact  团队影响

Top tips from the Inference team
推理团队的顶级建议

Accelerated ML concept learning
加速机器学习概念学习

Research time reduced by 80 % 80 % 80%80 \% - what took an hour of Google searching now takes 10-20 minutes.
研究时间减少了 80 % 80 % 80%80 \% ——原本需要一小时的 Google 搜索现在只需 10-20 分钟。

Faster codebase navigation
更快的代码库导航

Can find relevant files and understand system architecture in seconds instead of asking colleagues.
可以在几秒钟内找到相关文件并理解系统架构，而无需询问同事。

Comprehensive test coverage
全面的测试覆盖

Claude automatically generates unit tests with edge cases, relieving mental burden while maintaining code quality.
Claude 自动生成包含边界情况的单元测试，减轻心理负担，同时保持代码质量。

Language barrier elimination
消除语言障碍

Can implement functionality in unfamiliar languages like Rust without needing to learn it.
可以在不需要学习的情况下，用不熟悉的语言如 Rust 实现功能。

Test knowledge base functionality first
先测试知识库功能

Try asking various questions to see if Claude can answer faster than Google search. If it’s faster and more accurate, it’s a valuable time-saving tool for your workflow.
尝试提出各种问题，看看 Claude 是否能比 Google 搜索更快地回答。如果更快且更准确，它就是节省工作流程时间的宝贵工具。

Start with code generation
从代码生成开始

Give Claude specific instructions and ask it to write logic, then verify correctness. This helps build trust in the tool’s capabilities before using it for more complex tasks.
给 Claude 具体指令并让它编写逻辑，然后验证正确性。这有助于在使用它处理更复杂任务之前建立对该工具能力的信任。

Use it for test writing
用于编写测试

Having Claude write unit tests relieves significant pressure from daily development work. Leverage this feature to maintain code quality without spending time thinking through all test cases manually.
让 Claude 编写单元测试可以大大减轻日常开发工作的压力。利用此功能可以在不花时间手动思考所有测试用例的情况下保持代码质量。

Claude Code for data science and visualization
Claude Code 用于数据科学和可视化

Data Science and ML Engineering teams need sophisticated visualization tools to understand model performance, but building these tools often requires expertise in unfamiliar languages and frameworks.
数据科学和机器学习工程团队需要复杂的可视化工具来理解模型性能，但构建这些工具通常需要掌握不熟悉的语言和框架。
Claude Code enables these teams to build production-quality analytics dashboards without becoming full-stack developers.
Claude Code 使这些团队能够构建生产级的分析仪表盘，而无需成为全栈开发人员。

Main Claude Code use cases
Claude Code 主要使用场景

Building JavaScript/TypeScript dashboard apps
构建 JavaScript/TypeScript 仪表盘应用

Despite knowing “very little JavaScript and TypeScript,” the team uses Claude Code to build entire React applications for visualizing RL model performance and training data.
尽管“对 JavaScript 和 TypeScript 知之甚少”，团队仍然使用 Claude Code 构建完整的 React 应用程序，用于可视化 RL 模型性能和训练数据。
They give Claude control to write full applications from scratch, like a 5,000-line TypeScript app, without needing to understand the code themselves.
他们让 Claude 完全控制，从零开始编写完整的应用程序，比如一个 5000 行的 TypeScript 应用，而自己无需理解代码。
This is critical because visualization apps are relatively low context and don’t require understanding the entire monorepo, allowing rapid prototyping of tools to understand model performance during training and evaluations.
这非常关键，因为可视化应用上下文相对较少，不需要理解整个 monorepo，从而能够快速原型设计工具，以便在训练和评估过程中理解模型性能。

Handling repetitive refactoring tasks
处理重复的重构任务

When faced with merge conflicts or semi-complicated file refactoring that’s too complex for editor macros but not large enough for major development effort, they use Claude Code like a “slot machine” - commit their state, let Claude work autonomously for 30 minutes, and either accept the solution or restart fresh if it doesn’t work.
当遇到合并冲突或半复杂的文件重构，这些任务对编辑器宏来说过于复杂，但又不足以进行大规模开发时，他们将 Claude Code 当作“老虎机”使用——提交当前状态，让 Claude 自主工作 30 分钟，然后接受解决方案，或者如果不成功则重新开始。

Creating persistent analytics tools instead of throwaway notebooks
创建持久的分析工具，而非一次性笔记本

Instead of building one-off Jupyter notebooks that get discarded, the team now has Claude build permanent React dashboards that can be reused across future model evaluations.
团队不再构建一次性使用后丢弃的 Jupyter 笔记本，而是让 Claude 构建永久性的 React 仪表盘，可在未来的模型评估中重复使用。
This is important because understanding Claude’s performance is “one of the most important things for the team” - they need to understand how models perform during training and evaluations, which “is actually non-trivial and simple tools can’t get too much signal from looking at a single number go up.”
这很重要，因为了解 Claude 的性能是“团队最重要的事情之一”——他们需要了解模型在训练和评估过程中的表现，“这实际上并不简单，单靠一个数字的上升，简单的工具无法获得太多信号。”

Zero-dependency task delegation
零依赖任务委派

For tasks in completely unfamiliar codebases or languages, they delegate entire implementation to Claude Code, leveraging its ability to gather context from the monorepo and execute tasks without their involvement in the actual coding process.
对于完全不熟悉的代码库或语言中的任务，他们将整个实现委派给 Claude Code，利用其从单一代码库中收集上下文并执行任务的能力，而无需他们实际参与编码过程。
This allows productivity in areas outside their expertise instead of spending time learning new technologies.
这使得他们能够在自己不擅长的领域提高生产力，而不是花时间学习新技术。

Claude Code for data science and visualization
Claude Code 用于数据科学和可视化

Team impact  团队影响

Achieved 2-4x time savings
实现了 2-4 倍的时间节省

Routine refactoring tasks that were tedious but manageable manually are now completed much faster.
那些繁琐但可手动完成的常规重构任务现在完成得更快。

Built complex applications in unfamiliar languages
用不熟悉的语言构建复杂应用

Created 5,000-line TypeScript applications despite having minimal JavaScript/TypeScript experience.
尽管几乎没有 JavaScript/TypeScript 经验，仍然创建了 5000 行的 TypeScript 应用。

Shifted from throwaway to persistent tools
从一次性工具转向持久工具

Instead of disposable Jupyter notebooks, now building reusable React dashboards for model analysis.
不再使用一次性 Jupyter 笔记本，而是构建可重用的 React 仪表盘进行模型分析。

Direct model improvement insights
直接的模型改进见解

Firsthand Claude Code experience informs development of better memory systems and UX improvements for future model iterations.
第一手的 Claude Code 体验为未来模型迭代开发更好的记忆系统和用户体验改进提供了信息。

Enabled visualization-driven decision making
实现了基于可视化的决策制定

Better understanding of Claude’s performance during training and evaluations through advanced data visualization tools.
通过先进的数据可视化工具，更好地理解 Claude 在训练和评估过程中的表现。

Top tips from the Data Science and ML Engineering teams
数据科学和机器学习工程团队的顶级建议

Treat it like a slot machine
把它当作老虎机来对待

Save your state before letting Claude work, let it run for 30 minutes, then either accept the result or start fresh rather than trying to wrestle with corrections. Starting over often has a higher success rate than trying to fix Claude’s mistakes.
在让 Claude 工作之前保存你的状态，让它运行 30 分钟，然后要么接受结果，要么重新开始，而不是试图纠正错误。重新开始通常比修正 Claude 的错误成功率更高。

Interrupt for simplicity when needed
在需要时为了简化可以中断操作

While supervising, don’t hesitate to stop Claude and ask “why are you doing this? Try something simpler.” The model tends toward more complex solutions by default but responds well to requests for simpler approaches.
在监督时，不要犹豫，随时停止 Claude 并询问“你为什么这么做？试试更简单的方法。”该模型默认倾向于更复杂的解决方案，但对简化方法的请求反应良好。

Claude Code for API
Claude Code 的 API

The API Knowledge team works on features like PDF support, citations, and web search that bring additional knowledge into Claude’s context window.
API 知识团队致力于开发诸如 PDF 支持、引用和网页搜索等功能，将更多知识引入 Claude 的上下文窗口。
Working across large, complex codebases means constantly encountering unfamiliar code sections, spending significant time understanding which files to examine for any given task, and building context before making changes.
在处理大型复杂代码库时，意味着不断遇到不熟悉的代码部分，花费大量时间了解针对特定任务应查看哪些文件，并在进行更改前构建上下文。
Claude Code improves this experience by serving as a guide that can help them understand system architecture, identify relevant files, and explain complex interactions.
Claude Code 通过作为指导，帮助他们理解系统架构、识别相关文件并解释复杂的交互，从而提升了这一体验。

Main Claude Code use cases
Claude Code 的主要使用场景

First-step workflow planning
第一步工作流程规划

The team uses Claude Code as their “first stop” for any task, asking it to identify which files to examine for bug fixes, feature development, or analysis.
团队将 Claude Code 作为任何任务的“第一站”，让它帮助确定需要检查哪些文件以进行错误修复、功能开发或分析。
This replaces the traditional time-consuming process of manually navigating the codebase and gathering context before starting work.
这取代了传统的耗时过程，即在开始工作前手动浏览代码库并收集上下文信息。

Independent debugging across codebases
跨代码库的独立调试

The team now has the confidence to tackle bugs in unfamiliar parts of the codebase instead of asking others for help. They can ask Claude “Do you think you can fix this bug?
团队现在有信心处理代码库中不熟悉部分的错误，而不是寻求他人帮助。他们可以问 Claude：“你认为你能修复这个错误吗？
This is the behavior I’m seeing” and often get immediate progress, which wasn’t feasible before given the time investment required.
这是我观察到的行为”，并且通常能立即取得进展，而之前由于所需时间投入，这种情况是不可行的。

Model iteration testing through dogfooding
通过内部试用进行模型迭代测试

Claude Code automatically uses the latest research model snapshots, making it their primary way of experiencing model changes. This gives them direct feedback on model behavior changes during development cycles, which they hadn’t experienced during previous launches.
Claude Code 自动使用最新的研究模型快照，使其成为体验模型变化的主要方式。这让他们在开发周期中能够直接获得模型行为变化的反馈，这是以往发布中未曾经历过的。

Eliminating context-switching overhead
消除上下文切换的开销

Instead of copying code snippets and dragging files into Claude.ai while explaining problems extensively, they can ask questions directly in Claude Code without additional context gathering, significantly reducing mental overhead.
他们无需复制代码片段或拖拽文件到 Claude.ai 并进行大量问题说明，而是可以直接在 Claude Code 中提问，无需额外收集上下文，大大减少了认知负担。

Claude Code for API
Claude Code 用于 API

Team impact  团队影响

Top tips from the API Knowledge team
API 知识团队的顶级建议

Increased confidence in tackling unfamiliar areas
提升了应对陌生领域的信心

Team members can independently debug bugs and investigate incidents in unfamiliar codebases.
团队成员可以独立调试错误并调查不熟悉代码库中的事件。

Significant time savings in context gathering
在上下文收集方面节省了大量时间

Eliminated the overhead of copying code snippets and dragging files into Claude.ai, reducing mental context-switching burden.
消除了复制代码片段和将文件拖入 Claude.ai 的开销，减少了心理上下文切换的负担。

Faster rotation onboarding
更快的轮岗入职

Engineers rotating to new teams can quickly navigate unfamiliar codebases and contribute meaningfully without extensive colleague consultation.
轮换到新团队的工程师可以快速浏览不熟悉的代码库，并在无需大量同事咨询的情况下做出有意义的贡献。

Enhanced developer happiness
提升开发者满意度

Team reports feeling happier and more productive with reduced friction in daily workflows.
团队报告称，日常工作流程中的摩擦减少后，成员感到更快乐且更高效。

Treat it as an iterative partner, not a one-shot solution
将其视为一个迭代的合作伙伴，而非一次性解决方案。

Rather than expecting Claude to solve problems immediately, approach it as a collaborator you iterate with. This works better than trying to get perfect solutions on the first try.
不要期望 Claude 立即解决问题，而是将其视为一个可以反复迭代的协作者。这比试图第一次就获得完美解决方案效果更好。

Use it for building confidence in unfamiliar areas
用它来建立对不熟悉领域的信心

Don’t hesitate to tackle bugs or investigate incidents outside your expertise - Claude Code makes it feasible to work independently in areas that would normally require extensive context building.
不要犹豫去处理超出你专业范围的错误或调查事件——Claude Code 使得在通常需要大量上下文构建的领域中独立工作成为可能。

Start with minimal information
从最少的信息开始

Begin with just the bare minimum of what you need and let Claude guide you through the process, rather than front-loading extensive explanations.
从你所需的最基本内容开始，让 Claude 引导你完成整个过程，而不是一开始就提供大量详尽的解释。

Claude Code for growth marketing
增长营销的 Claude Code

The Growth Marketing team focuses on building out performance marketing channels across paid search, paid social, mobile app stores, email marketing, and SEO.
增长营销团队专注于构建涵盖付费搜索、付费社交、移动应用商店、电子邮件营销和 SEO 的绩效营销渠道。
As a nontechnical team of one, they use Claude Code to automate repetitive marketing tasks and create agentic workflows that would traditionally require significant engineering resources.
作为一个非技术人员组成的单人团队，他们使用 Claude Code 自动化重复的营销任务，并创建传统上需要大量工程资源的自主工作流程。

Main Claude Code use cases
主要的 Claude Code 使用案例

Automated Google Ads creative generation
自动化 Google Ads 创意生成

The team built an agentic workflow that processes CSV files containing hundreds of existing ads with performance metrics, identifies underperforming ads for iteration, and generates new variations that meet strict character limits ( 30 characters for headlines, 90 for descriptions).
团队构建了一个代理工作流，处理包含数百个带有性能指标的现有广告的 CSV 文件，识别表现不佳的广告进行迭代，并生成符合严格字符限制的新变体（标题限 30 字符，描述限 90 字符）。
Using two specialized sub-agents (one for headlines, one for descriptions), the system can generate hundreds of new ads in minutes instead of requiring manual creation across multiple campaigns.
通过使用两个专门的子代理（一个负责标题，一个负责描述），系统可以在几分钟内生成数百个新广告，而无需在多个广告系列中手动创建。
This has enabled them to test and iterate at scale, something that would have taken a significant amount of time to achieve previously.
这使他们能够大规模测试和迭代，而这在以前需要花费大量时间才能实现。

Figma plugin for mass creative production
用于大规模创意制作的 Figma 插件

Instead of manually duplicating and editing static images for paid social ads, they developed a Figma plugin that identifies frames and programmatically generates up to 100 ad variations by swapping headlines and descriptions, reducing what would take hours of copypasting to half a second per batch.
他们开发了一个 Figma 插件，能够识别画框并通过程序化方式生成多达 100 个广告变体，替换标题和描述，取代了手动复制和编辑付费社交广告中的静态图像，将需要数小时的复制粘贴缩短到每批次半秒。
This enables 10x creative output, allowing the team to test vastly more creative variations across key social channels.
这使创意产出提升了 10 倍，团队能够在主要社交渠道上测试更多的创意变体。

Meta Ads MCP server for campaign analytics
Meta Ads MCP 服务器用于活动分析

They created an MCP server integrated with Meta Ads API to query campaign performance, spending data, and ad effectiveness directly within the Claude Desktop app, eliminating the need to switch between platforms for performance analysis, saving critical time where every efficiency gain translates to better ROI.
他们创建了一个集成了 Meta Ads API 的 MCP 服务器，可以直接在 Claude Desktop 应用中查询活动表现、支出数据和广告效果，免去了在不同平台间切换进行表现分析的需要，节省了关键时间，每一点效率提升都转化为更好的投资回报率。

Advanced prompt engineering with memory systems
带有记忆系统的高级提示工程

They implemented a rudimentary memory system that logs hypotheses and experiments across ad iterations, allowing the system to pull previous test results into context when generating new variations, creating a selfimproving testing framework.
他们实现了一个初步的记忆系统，记录广告迭代中的假设和实验，使系统在生成新变体时能够将之前的测试结果纳入上下文，创建了一个自我改进的测试框架。
This enables systematic experimentation that would be impossible to track manually.
这使得系统化实验成为可能，而手动跟踪则不可能实现。

Claude Code for growth marketing
用于增长营销的 Claude Code

Team impact  团队影响

Dramatic time savings on repetitive tasks
在重复性任务上显著节省时间

Ad copy creation reduced from 2 hours to 15 minutes, freeing up time for strategic work.
广告文案创作时间从 2 小时缩短到 15 分钟，释放出更多时间用于战略工作。

10x increase in creative output
创意产出提升 10 倍

The team can now test vastly more ad variations across channels with automated generation and Figma integration.
团队现在可以通过自动生成和 Figma 集成，在各渠道测试更多广告变体。

Operating like a larger team
运作如同一个更大的团队

The team can handle tasks that traditionally required dedicated engineering resources.
团队能够处理传统上需要专门工程资源的任务。

Strategic focus shift  战略重点转移

The team can spend more time on overall strategy and building agentic automation rather than manual execution.
团队可以将更多时间用于整体战略和构建自主自动化，而非手动执行。

Top tips from the Growth Marketing team
增长营销团队的顶级建议

Identify API-enabled repetitive tasks
识别支持 API 的重复任务

Look for workflows involving repetitive actions with tools that have APIs (like ad platforms, design tools, analytics platforms). These are prime candidates for automation and where Claude Code provides the most value.
寻找涉及使用带有 API 的工具（如广告平台、设计工具、分析平台）执行重复操作的工作流程。这些是自动化的最佳候选场景，也是 Claude Code 提供最大价值的地方。

Break complex workflows into specialized sub-agents
将复杂工作流程拆分为专门的子代理

Instead of trying to handle everything in one prompt or workflow, create separate agents for specific tasks (like their headline agent vs. description agent). This makes debugging easier and improves output quality when dealing with complex requirements.
不要试图在一个提示或工作流程中处理所有内容，而是为特定任务创建独立的代理（例如他们的标题代理与描述代理）。这使调试更容易，并在处理复杂需求时提升输出质量。

Thoroughly brainstorm and prompt plan before coding
在编码前彻底进行头脑风暴和提示规划

Spend significant time upfront using Claude.ai to think through your entire workflow, then have Claude.ai create a comprehensive prompt and code structure for Claude Code to reference. Also, work step-by-step rather than asking for one-shot solutions to avoid Claude getting overwhelmed by complex tasks.
花大量时间使用 Claude.ai 预先思考整个工作流程，然后让 Claude.ai 创建一个全面的提示和代码结构供 Claude Code 参考。同时，采用逐步进行的方法，而不是一次性请求解决方案，以避免 Claude 被复杂任务压垮。

Claude Code for product design
用于产品设计的 Claude Code

The Product Design team supports Claude Code, Claude.ai and the Anthropic API, specializing in building AI products. Even non-developers can use Claude Code to bridge the traditional gap between design and engineering, enabling direct implementation of their design vision without extensive back-and-forth with engineers.
产品设计团队支持 Claude Code、Claude.ai 和 Anthropic API，专注于构建 AI 产品。即使是非开发人员也可以使用 Claude Code 弥合设计与工程之间的传统鸿沟，实现设计愿景的直接落地，无需与工程师反复沟通。

Main Claude Code use cases
Claude Code 的主要使用场景

Front-end polish and state management changes
前端优化和状态管理变更

Instead of creating extensive design documentation and going through multiple rounds of feedback with engineers for visual tweaks (typefaces, colors, spacing), they now directly implement these changes using Claude Code.
他们不再创建大量设计文档，也不需要与工程师反复沟通视觉调整（字体、颜色、间距），而是直接使用 Claude Code 实现这些更改。
Engineers noted they’re making “large state management changes that you typically wouldn’t see a designer making,” enabling them to achieve the exact quality they envision.
工程师们指出，他们正在进行“通常设计师不会做的大规模状态管理变更”，这使他们能够实现理想中的精确质量。

GitHub Actions automated ticketing
GitHub Actions 自动工单

Using Claude Code’s GitHub integration, they can simply file issues/ tickets describing needed changes, and Claude automatically proposes code solutions without having to open Claude Code, creating a seamless bug-fixing and feature refinement workflow for their persistent backlog of polish tasks.
通过 Claude Code 的 GitHub 集成，他们可以轻松提交描述所需更改的问题/工单，Claude 会自动提出代码解决方案，无需打开 Claude Code，从而为他们持续积压的优化任务创建了无缝的修复和功能完善工作流程。

Rapid interactive prototyping
快速交互式原型设计

By pasting mockup images into Claude Code, they generate fully functional prototypes that engineers can immediately understand and iterate on, replacing the traditional cycle of static Figma designs that required extensive explanation and translation to working code.
通过将原型图像粘贴到 Claude Code 中，他们生成了工程师可以立即理解和迭代的全功能原型，取代了传统的静态 Figma 设计周期，这些设计通常需要大量解释和转换为可运行代码。

Edge case discovery and system architecture understanding
边缘情况发现与系统架构理解

They use Claude Code to map out error states, logic flows, and different system statuses, allowing them to identify edge cases during design rather than discovering them later in development, fundamentally improving the quality of their initial designs.
他们使用 Claude Code 绘制错误状态、逻辑流程和不同的系统状态图，使他们能够在设计阶段识别边缘情况，而不是在开发后期才发现，从根本上提升了初始设计的质量。
For tasks like removing “research preview” messaging across the entire codebase, they used Claude Code to find all instances, review surrounding copy, coordinate changes with legal in real-time, and implement updates a process that took two 30-minute calls instead of a week of back-andforth coordination.
对于在整个代码库中移除“研究预览”提示等任务，他们使用 Claude Code 查找所有实例，审查周围文案，实时与法律部门协调变更，并实施更新，这一过程通过两次各 30 分钟的电话完成，而不是一周的反复协调。

Claude Code for product design
Claude Code 用于产品设计

Team impact  团队影响

Transformed core workflow
核心工作流程的转变

Claude Code becomes a primary design tool, with Figma and Claude Code open 80 % 80 % 80%80 \% of the time.
Claude Code 成为主要设计工具，Figma 和 Claude Code 同时打开的时间占 80 % 80 % 80%80 \% 。

2-3x faster execution  执行速度提升 2-3 倍

Visual and state management changes that previously required extensive back-and-forth with engineers now implemented directly.
以前需要与工程师反复沟通的视觉和状态管理更改现在直接实现。

Weeks to hours cycle time
周期时间从数周缩短到数小时

Complex projects like GA launch messaging that would take a week of coordination now completed in two 30-minute calls.
像 GA 发布消息这样复杂的项目，原本需要一周的协调，现在通过两次 30 分钟的电话会议完成。

Two distinct user experiences
两种截然不同的用户体验

Developers get “augmented workflow” (faster execution), while nontechnical users get “holy crap, I’m a developer workflow” (entirely new capabilities previously impossible).
开发者获得“增强工作流”（更快的执行速度），而非技术用户则获得“天哪，我成了开发者的工作流”（全新能力，以前不可能实现）。

Improved design-engineering collaboration
改进的设计与工程协作

Better communication and faster problem-solving because designers understand system constraints and possibilities upfront.
更好的沟通和更快的问题解决，因为设计师能够提前了解系统的限制和可能性。

Top tips from the Product Design team
产品设计团队的顶级建议

Get proper setup help from engineers
寻求工程师的正确设置帮助

Have engineering teammates help with initial repository setup and permissions - the technical onboarding is challenging for non-developers, but once configured, it becomes transformative for daily workflow.
让工程团队成员协助完成初始代码仓库设置和权限配置——技术入职对非开发人员来说具有挑战性，但一旦配置完成，将极大改善日常工作流程。

Use custom memory files to guide Claude's behavior
使用自定义记忆文件来引导 Claude 的行为

Create specific instructions telling Claude you’re a designer with little coding experience who needs detailed explanations and smaller, incremental changes, dramatically improving the quality of Claude’s responses and making it less intimidating.
创建具体的指令，告诉 Claude 你是一名编码经验有限的设计师，需要详细的解释和较小的渐进式更改，这大大提升了 Claude 的响应质量，并使其不那么令人生畏。

Leverage image pasting for prototyping
利用图像粘贴进行原型设计

Use Command +V to paste screenshots directly into Claude Code - it excels at reading designs and generating functional code, making it invaluable for turning static mockups into interactive prototypes that engineers can immediately understand and build upon.
使用 Command + V 直接将截图粘贴到 Claude Code 中——它擅长读取设计并生成功能代码，对于将静态模型转换为工程师可以立即理解和构建的交互式原型非常宝贵。

Claude Code for RL engineering
Claude Code 用于强化学习工程

The RL Engineering team focuses on efficient sampling in RL and weight transfers across the cluster.
RL 工程团队专注于强化学习中的高效采样和集群间的权重传递。
They use Claude Code primarily for writing small to medium features, debugging, and understanding complex codebases, with an iterative approach that includes frequent checkpointing and rollbacks.
他们主要使用 Claude Code 来编写中小型功能、调试以及理解复杂代码库，采用包括频繁检查点和回滚的迭代方法。

Main Claude Code use cases
Claude Code 主要使用场景

Feature development with supervised autonomy
带有监督自治的功能开发

The team lets Claude Code write most of the code for small to medium features while providing oversight, such as implementing authentication mechanisms for weight transfer components.
团队让 Claude Code 为小到中等规模的功能编写大部分代码，同时提供监督，例如为权重传输组件实现认证机制。
They work interactively, allowing Claude to take the lead but steering it when it goes off track.
他们以交互方式工作，允许 Claude 主导，但在其偏离方向时进行引导。

Test generation and code review
测试生成和代码审查

After implementing changes themselves, they ask Claude Code to add tests or review their code. This automated testing workflow saves significant time on routine but important quality assurance tasks.
在自己实现更改后，他们会让 Claude Code 添加测试或审查代码。这种自动化测试工作流程在常规但重要的质量保证任务上节省了大量时间。

Debugging and error investigation
调试和错误调查

They use Claude Code to debug errors with mixed results - sometimes it identifies issues immediately and adds relevant tests, while other times it struggles to understand the problem, but overall provides value when it works.
他们使用 Claude Code 进行错误调试，效果参差不齐——有时它能立即识别问题并添加相关测试，另一些时候则难以理解问题，但总体上在有效时能提供价值。

Codebase comprehension and call stack analysis
代码库理解和调用栈分析

One of the biggest changes in their workflow is using Claude Code to get quick summaries of relevant components and call stacks, replacing manual code reading or extensive debugging output generation.
他们工作流程中最大的变化之一是使用 Claude Code 快速获取相关组件和调用栈的摘要，取代了手动阅读代码或生成大量调试输出。

Kubernetes operations guidance
Kubernetes 操作指南

They frequently ask Claude Code about Kubernetes operations that would otherwise require extensive Googling, getting immediate answers for configuration and deployment questions.
他们经常向 Claude Code 询问 Kubernetes 操作，这些操作本来需要大量谷歌搜索，从而能立即获得配置和部署问题的答案。

Claude Code for RL engineering
用于强化学习工程的 Claude Code

Development workflow impact
开发工作流程的影响

Experimental approach enabled
启用实验性方法

They now use a “try and rollback” methodology, frequently committing checkpoints so they can test Claude’s autonomous implementation attempts and revert if needed, enabling more experimental.
他们现在采用“尝试并回滚”的方法，频繁提交检查点，以便测试 Claude 的自主实现尝试，并在需要时回退，从而支持更多实验性操作。

Documentation acceleration
文档加速

Claude Code automatically adds helpful comments that save significant time on documentation, though they note it sometimes adds comments in odd places or uses questionable code organization.
Claude Code 会自动添加有用的注释，大大节省了文档编写时间，尽管他们指出有时注释会出现在奇怪的位置或代码组织存在问题。

Speed-up with limitations
加速但有限制

While Claude Code can implement small-to-medium PRs with “relatively little time” from them, they acknowledge it only works on first attempt about one-third of the time, requiring either additional guidance or manual intervention.
虽然 Claude Code 可以在“相对较短的时间”内实现小到中等规模的 PR，但他们承认它只有大约三分之一的概率在第一次尝试时成功，通常需要额外的指导或人工干预。

Top tips from the RL Engineering team
RL 工程团队的顶级建议

Customize your Claude.md file for specific patterns
为特定模式自定义你的 Claude.md 文件

Add instructions to your Claude.md file to prevent Claude from making repeated tool-calling mistakes, such as telling it to “run pytest not run and don’t cd unnecessarily - just use the right path.” This significantly improved consistency.
在你的 Claude.md 文件中添加指令，防止 Claude 重复调用工具时出错，比如告诉它“运行 pytest，不要运行其他命令，也不要不必要地切换目录——只需使用正确的路径。”这大大提高了一致性。

Use a checkpoint-heavy workflow
使用以检查点为主的工作流程

Regularly commit your work as Claude makes changes so you can easily roll back when experiments don’t work out. This enables a more experimental approach to development without risk.
随着 Claude 的更改，定期提交你的工作，这样当实验失败时可以轻松回滚。这使得开发可以采取更具实验性的方式而无风险。

Try one-shot first, then collaborate
先尝试一次性完成，然后再进行协作

Give Claude a quick prompt and let it attempt the full implementation first. If it works (about one-third of the time), you’ve saved significant time. If not, then switch to a more collaborative, guided approach.
先给 Claude 一个简短的提示，让它先尝试完整实现。如果成功了（大约三分之一的概率），你就节省了大量时间。如果不成功，则切换到更协作、引导式的方法。
The Legal team discovered Claude Code’s potential through experimentation, and a desire to learn about Anthropic’s product offerings.
法律团队通过实验发现了 Claude Code 的潜力，并希望了解 Anthropic 的产品。
Additionally, one team member had a personal use case related to creating accessibility tools for family and work prototypes that demonstrate the technology’s power for non-developers.
此外，一名团队成员有一个个人用例，涉及为家庭和工作原型创建辅助工具，展示了该技术对非开发者的强大作用。

Main Claude Code use cases
Claude Code 主要使用案例

Custom accessibility solution for family members
为家庭成员定制的无障碍解决方案

Team members have built communication assistants for family members with speaking difficulties due to medical diagnoses.
团队成员为因医疗诊断导致言语障碍的家庭成员开发了沟通助手。
In just one hour, they created a predictive text app using native speech-to-text that suggests responses and speaks them using voice banks, solving gaps in existing accessibility tools recommended by speech therapists.
仅用一小时，他们利用原生语音转文本技术创建了一个预测文本应用，能够建议回复内容并通过语音库朗读，解决了言语治疗师推荐的现有无障碍工具中的不足。
They created prototype “phone tree” systems to help team members connect with the right lawyer at Anthropic, demonstrating how legal departments can build custom tools for common tasks without traditional development resources.
他们创建了原型“电话树”系统，帮助团队成员联系 Anthropic 的合适律师，展示了法律部门如何在没有传统开发资源的情况下，为常见任务构建定制工具。

Team coordination tools  团队协调工具

Managers have built G Suite applications that automate weekly team updates and track legal review status across products, allowing lawyers to quickly flag items needing review through simple button clicks rather than spreadsheet management.
经理们构建了 G Suite 应用程序，自动化每周团队更新并跟踪各产品的法律审查状态，使律师能够通过简单的按钮点击快速标记需要审查的事项，而无需管理电子表格。

Rapid prototyping for solution validation
快速原型制作以验证解决方案

They use Claude Code to quickly build functional prototypes they can show to domain experts (like showing accessibility tools to UCSF specialists) to validate ideas and identify existing solutions before investing more time.
他们使用 Claude Code 快速构建功能原型，展示给领域专家（例如向 UCSF 专家展示无障碍工具），以验证想法并在投入更多时间之前识别现有解决方案。

Work style and impact
工作风格与影响

Security and compliance awareness
安全与合规意识

Planning in Claude.ai, building in Claude Code
在 Claude.ai 中进行规划，在 Claude Code 中构建

They use a two-step process where they brainstorm and plan with Claude.ai first, then move to Claude Code for implementation, asking it to slow down and work step-by-step rather than outputting everything at once.
他们采用两步法，先在 Claude.ai 中进行头脑风暴和规划，然后转到 Claude Code 进行实现，要求它放慢速度，逐步工作，而不是一次性输出所有内容。

Visual-first approach  以视觉为先的方法

They frequently use screenshots to show Claude Code what they want interfaces to look like, then iterate based on visual feedback rather than describing features in text.
他们经常使用截图向 Claude Code 展示他们希望界面呈现的样子，然后根据视觉反馈进行迭代，而不是通过文字描述功能。

Prototype-driven innovation
原型驱动的创新

They emphasize overcoming the fear of sharing “silly” or “toy” prototypes, as these demonstrations inspire others to see possibilities they hadn’t considered.
他们强调克服分享“傻瓜”或“玩具”原型的恐惧，因为这些演示能激发他人看到他们未曾考虑过的可能性。

MCP integration concerns
MCP 集成问题

As product lawyers, they immediately identify security implications of deep MCP integrations, noting how conservative security postures will create barriers as AI tools access more sensitive systems.
作为产品律师，他们立即识别出深度 MCP 集成的安全影响，指出随着 AI 工具访问更敏感的系统，保守的安全态度将会形成障碍。

Compliance tooling priorities
合规工具优先级

They advocate for building compliance tools quickly as AI capabilities expand, recognizing the balance between innovation and risk management.
他们主张随着 AI 能力的扩展，快速构建合规工具，认识到创新与风险管理之间的平衡。

Plan extensively in Claude.ai first
首先在 Claude.ai 中进行广泛规划

Use Claude’s conversational interface to flesh out your entire idea before moving to Claude Code. Then ask Claude to summarize everything into a step-by-step prompt for implementation.
使用 Claude 的对话界面来完善您的整个想法，然后再转到 Claude Code。接着让 Claude 将所有内容总结成逐步的实现提示。

Work incrementally and visually
逐步且可视化地工作

Ask Claude Code to slow down and implement one step at a time so you can copy-paste without getting overwhelmed. Use screenshots liberally to show what you want interfaces to look like.
让 Claude Code 放慢速度，一次实现一个步骤，这样您可以复制粘贴而不会感到不知所措。大量使用截图来展示您希望界面呈现的样子。

Share prototypes despite imperfection
即使不完美，也要分享原型

Overcome the urge to hide “toy” projects or unfinished work - sharing prototypes helps others see possibilities and sparks innovation across departments that don’t typically interact.
克服隐藏“玩具”项目或未完成工作的冲动——分享原型有助于他人看到可能性，并激发通常不互动的部门之间的创新。
Al