How Anthropic teams use Claude Code
Anthropic 團隊如何運用 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 的內部團隊正透過 Claude Code 改變工作流程，讓開發人員和非技術人員都能處理複雜專案、自動化任務，並彌補過去限制生產力的技能差距。

To learn more, we spoke with the following teams:
為了進一步了解，我們與以下團隊進行了交流：

• Data infrastructure  資料基礎架構
• Product development  產品開發
• Security engineering  安全工程
• Inference  推論
• Data science and visualization
  資料科學與視覺化
• Product engineering  產品工程
• Growth marketing  成長行銷
• Product design  產品設計
• Reinforcement learning (RL) engineering
  強化學習（RL）工程
• Legal  法務部門

Through these interviews, we've gathered insights on how different departments use Claude Code, its impact on their work, and tips for other organizations considering adoption.
透過這些訪談，我們收集了不同部門如何使用 Claude Code 的見解、對他們工作的影響，以及考慮採用的其他組織的建議。

Claude Code for data infrastructure
Claude Code 用於數據基礎架構

The Data Infrastructure team organizes all business data for teams across the company. 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.
數據基礎架構團隊負責整理全公司各團隊的業務數據。他們使用 Claude Code 來自動化例行數據工程任務、疑難排解複雜的基礎架構問題，並建立有文件記錄的工作流程，讓技術與非技術團隊成員都能獨立存取和操作數據。

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

Kubernetes debugging with screenshots
搭配螢幕截圖進行 Kubernetes 除錯

When Kubernetes clusters went down and weren't scheduling new pods, the team used Claude Code to diagnose the issue. 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. 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.
當 Kubernetes 叢集發生故障且無法調度新 pod 時，團隊使用 Claude Code 來診斷問題。他們將儀表板的截圖輸入 Claude Code，該工具逐步引導他們瀏覽 Google Cloud 的 UI 選單，直到發現一個警告顯示 pod IP 地址耗盡。接著 Claude Code 提供了確切的指令來創建新 IP 池並將其加入叢集，省去了需要網路專家介入的麻煩。

Plain text workflows for finance team
財務團隊的純文字工作流程

Engineers 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. 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.
工程師向財務團隊成員展示如何撰寫描述其數據工作流程的純文字檔案，然後將其載入 Claude Code 以獲得全自動執行。沒有編碼經驗的員工可以描述步驟如「查詢此儀表板、獲取資訊、執行這些查詢、產生 Excel 輸出」，而 Claude Code 會執行整個工作流程，包括詢問所需的輸入如日期。

Codebase navigation for new hires
新進人員的程式碼庫導覽

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.
當新的資料科學家加入團隊時，他們會被引導使用 Claude Code 來瀏覽龐大的程式碼庫。Claude Code 會讀取他們的 Claude.md 文件（說明文件），識別特定任務的相關檔案，解釋資料管線的相依性，並幫助新人理解哪些上游資料來源會匯入儀表板。這取代了傳統的資料目錄和探索工具。

End-of-session documentation updates
工作階段結束時的說明文件更新

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.
團隊會要求 Claude Code 總結已完成的工作階段，並在每項任務結束時提出改進建議。這形成了一個持續改進的循環，讓 Claude Code 能根據實際使用情況來精煉 Claude.md 說明文件和工作流程指引，使後續的迭代更加有效。

Parallel task management across multiple instances
跨多個執行個體的平行任務管理

When working on long-running data tasks, the team opens multiple instances of Claude Code in different repositories for different projects. 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.
在處理長時間運行的數據任務時，團隊會針對不同專案在不同程式庫中開啟多個 Claude Code 實例。每個實例都能保持完整的上下文，因此當他們在數小時或數天後切換回來時，Claude Code 能準確記住他們之前的工作進度和中斷點，實現真正的平行工作流程管理而不會遺失上下文。

Team impact  團隊影響力

Resolved infrastructure problems without specialized expertise
無需專業知識即解決基礎架構問題

Resolves Kubernetes cluster issues that would normally require pulling in systems or networking team members, using Claude Code to diagnose problems and provide exact fixes.
運用 Claude Code 診斷問題並提供精確修正方案，解決了通常需要系統或網路團隊成員介入的 Kubernetes 叢集問題。

Accelerated onboarding  加速上手

New data analysts and team members can quickly understand complex systems and contribute meaningfully without extensive guidance.
新進的數據分析師和團隊成員能夠快速理解複雜系統，無需過多指導即可做出有意義的貢獻。

Enhanced support workflow
強化支援流程

Claude Code can process much larger data volumes and identify anomalies (like monitoring 200 dashboards) that would be impossible for humans to review manually.
Claude Code 能夠處理更龐大的數據量，並識別出人類無法手動審查的異常情況（例如監控 200 個儀表板）。

Enabled cross-team self-service
啟用跨團隊自助服務

Finance teams with no coding experience can now execute complex data workflows independently.
沒有程式經驗的財務團隊現在可以獨立執行複雜的資料工作流程

Top tips from the Data Infrastructure team
來自資料基礎設施團隊的頂尖建議

Write detailed Claude.md files
撰寫詳細的 Claude.md 檔案

According to the team the better you document your workflows, tools, and expectations in Claude.md files, the better Claude Code performs. This makes Claude Code excel at routine tasks like setting up new data pipelines when you have existing design patterns.
根據團隊的說法，你在 Claude.md 文件中記錄的工作流程、工具和期望越詳細，Claude Code 的表現就會越好。這讓 Claude Code 在處理常規任務時表現出色，例如當你已有現成的設計模式時，它能快速建立新的數據管道。

Use MCP servers instead of CLI for sensitive data
對於敏感數據，請使用 MCP 伺服器而非 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.
他們建議使用 MCP 伺服器而非 BigQuery CLI，以便更好地控制 Claude Code 能存取的內容，特別是在處理需要記錄或可能涉及隱私問題的敏感數據時。

Share team usage sessions
分享團隊使用時段

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.
團隊舉辦了交流會議，成員們互相展示各自的 Claude Code 工作流程。這有助於推廣最佳實踐，並展現出他們可能獨自使用時未曾發現的不同工具應用方式。

Claude Code for product development
產品開發中的 Claude Code 應用

The Claude Code Product Development team uses their own product to build updates to Claude Code, expanding the product's enterprise capabilities and agentic loop functionalities.
Claude Code 產品開發團隊運用自家產品來建構 Claude Code 的更新版本，持續擴展產品的企業級功能與智能代理循環功能。

Main Claude Code use cases
Claude Code 主要應用場景

Fast prototyping with auto-accept mode
快速原型設計與自動接受模式

Engineers use Claude Code for rapid prototyping by enabling "auto-accept mode" (shift+tab) and setting up autonomous loops in which Claude writes code, runs tests, and iterates continuously. They give Claude abstract problems they're unfamiliar with, let it work autonomously, then review the 80% complete solution before taking over for final refinements. The team suggests starting from a clean git state and committing checkpoints regularly so they can easily revert any incorrect changes if Claude goes off track.
工程師們使用 Claude Code 進行快速原型設計，方法是啟用「自動接受模式」(shift+tab)並設置自主循環，讓 Claude 自動編寫代碼、運行測試並持續迭代。他們會給 Claude 一些自己不熟悉的抽象問題，讓它自主工作，然後在接手進行最終完善前，先審查已完成 80%的解決方案。團隊建議從乾淨的 git 狀態開始，並定期提交檢查點，這樣當 Claude 偏離軌道時，就能輕鬆還原任何錯誤的更改。

Synchronous coding for core features
核心功能的同步編碼

For more critical features touching the application's business logic, the team works synchronously with Claude Code, giving detailed prompts with specific implementation instructions. 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.
對於涉及應用程式業務邏輯的更關鍵功能，團隊會與 Claude Code 同步工作，提供包含具體實現指令的詳細提示。他們會即時監控整個過程，以確保代碼品質、符合風格指南和正確的架構，同時讓 Claude 處理重複性的編碼工作。

Building Vim mode  打造 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, and roughly 70% of the final implementation came from Claude's autonomous work, requiring only a few iterations to complete.
他們最成功的非同步專案之一，就是為 Claude Code 實作 Vim 鍵盤綁定功能。團隊直接請 Claude 建構整個功能，最終約有 70% 的實作內容來自 Claude 自主完成，僅需少量迭代即可完工。

Test generation and bug fixes
測試生成與錯誤修復

The team uses Claude Code to write comprehensive tests after implementing features and handle simple bug fixes identified in pull request reviews. They also use GitHub Actions to have Claude automatically address Pull Request comments like formatting issues or function renaming.
團隊會運用 Claude Code 在功能實作後撰寫完整測試，並處理程式碼審查中發現的簡單錯誤修復。他們也透過 GitHub Actions 讓 Claude 自動處理 Pull Request 中的註解，像是格式問題或函式重新命名等事項。

Codebase exploration  程式碼庫探索

When working with unfamiliar codebases (like the monorepo or API side), the team uses Claude Code to quickly understand how systems work. Instead of waiting for Slack responses, they ask Claude directly for explanations and code references, saving significant time in context switching.
當團隊需要處理不熟悉的程式碼庫（例如單體式倉庫或 API 端）時，他們會使用 Claude Code 來快速理解系統運作方式。與其等待 Slack 上的回覆，團隊直接向 Claude 詢問解釋和程式碼參考，大幅節省了上下文切換所需的時間。

Team impact  團隊影響

Faster feature implementation
更快速的功能實現

Claude Code successfully implemented complex features like Vim mode with 70% of code written autonomously by Claude.
Claude Code 成功實現了如 Vim 模式等複雜功能，其中 70%的代碼由 Claude 自主完成

Improved development velocity
提升開發速度

The tool rapidly prototypes features and iterates on ideas without getting bogged down in implementation details.
該工具能快速原型化功能並迭代想法，不會陷入實作細節的泥沼

Enhanced code quality through automated testing
透過自動化測試提升代碼品質

Claude generates comprehensive tests and handles routine bug fixes, maintaining high standards while reducing manual effort.
Claude 能生成全面的測試並處理例行錯誤修復，在維持高標準的同時減少人工投入。

Better codebase exploration
更完善的程式碼庫探索

Team members quickly get up to speed on unfamiliar parts of the monorepo without waiting for colleague responses.
團隊成員能快速熟悉 monorepo 中不熟悉的部分，無需等待同事回覆。

Top tips from the Claude Code team
Claude Code 團隊的頂尖秘訣

Create self-sufficient loops
建立自我完善的循環

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.
設定 Claude 自動執行建構、測試和程式碼檢查來驗證自身工作。這讓 Claude 能更長時間自主運作並自行發現錯誤，特別是在要求 Claude 先寫測試再寫程式碼時效果更佳。

Develop task classification intuition
培養任務分類的直覺

Learn to distinguish between tasks that work well asynchronously (peripheral features, prototyping) versus those needing synchronous supervision (core business logic, critical fixes). Abstract tasks on the product's edges can be handled with "auto-accept mode," while core functionality requires closer oversight.
學會區分適合非同步處理的任務（周邊功能、原型開發）與需要同步監督的任務（核心業務邏輯、關鍵修復）。產品邊緣的抽象任務可使用「自動接受模式」處理，而核心功能則需要更密切的監督。

Form clear, detailed prompts
撰寫清晰、詳細的提示

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.
當元件名稱或功能相似時，請在請求中特別明確標示。您的提示越完善、越詳細，就越能信任 Claude 獨立作業，而不會意外更動到程式庫中錯誤的部分。

Claude Code for security engineering
Claude Code 在資安工程中的應用

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.
資安工程團隊專注於確保軟體開發生命週期的安全性、供應鏈安全及開發環境安全。他們廣泛使用 Claude Code 來編寫和除錯程式碼。

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

Complex infrastructure debugging
複雜基礎設施除錯

When working on incidents, they feed Claude Code stack traces and documentation, asking it to trace control flow through the codebase. 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.
在處理事故時，他們會將堆疊追蹤和文件輸入 Claude Code，要求它追蹤程式碼庫中的控制流程。這大幅縮短了生產問題的解決時間，讓他們能在約 5 分鐘內理解通常需要 10-15 分鐘手動掃描程式碼才能發現的問題。

Terraform code review and analysis
Terraform 程式碼審查與分析

For infrastructure changes requiring security approval, the team copies Terraform plans into Claude Code to ask "what's this going to do? 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.
對於需要安全審核的基礎架構變更，團隊會將 Terraform 計畫複製到 Claude Code 中詢問「這會做什麼？我會後悔嗎？」這樣能建立更緊密的回饋循環，讓安全團隊能更快速地審查並核准基礎架構變更，減少開發流程中的瓶頸。

Documentation synthesis and runbooks
文件彙整與操作手冊

Claude Code ingests multiple documentation sources and creates markdown runbooks, troubleshooting guides, and overviews. The team uses these condensed documents as context for debugging real issues, creating a more efficient workflow than searching through full knowledge bases.
Claude Code 能整合多份文件來源，自動生成 Markdown 格式的操作手冊、故障排除指南與系統概覽。團隊將這些精簡文件作為除錯時的參考依據，比起翻閱完整知識庫，大幅提升了工作流程效率。

Test-driven development workflow
測試驅動開發流程

Instead of their previous "design doc → janky code → refactor → 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.
團隊捨棄過去「設計文件→草率編碼→重構→放棄測試」的模式，現在會先向 Claude Code 索取虛擬碼，引導其進行測試驅動開發，並在卡關時定期檢查調整方向，最終產出更可靠且具可測試性的程式碼。

Context switching and project onboarding
上下文切換與專案導入

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.
在參與現有專案（如「dependant」——一個用於安全審核流程的網路應用程式）時，他們使用 Claude Code 來撰寫、審查並執行以 markdown 格式撰寫並儲存在程式碼庫中的規格說明，使得能在數日內而非數週內做出實質貢獻。

Team impact  團隊影響

Reduced incident resolution time
縮短事件解決時間

Infrastructure debugging that normally takes 10-15 minutes of manual code scanning now takes about 5 minutes.
原本需要 10-15 分鐘手動掃描程式碼的基礎設施除錯工作，現在只需約 5 分鐘即可完成。

Improved security review cycle
提升安全審查效率

Terraform code reviews for security approval happen much faster, eliminating developer blocks while waiting for security team approval.
Terraform 程式碼的安全審查核准速度大幅加快，消除了開發人員等待安全團隊核准時的阻滯。

Enhanced cross-functional contribution
強化跨職能協作貢獻

Team members can meaningfully contribute to projects within days instead of weeks of context building.
團隊成員能在數日內對專案做出實質貢獻，無需耗費數週時間建立背景知識。

Better documentation workflow
更完善的文檔工作流程

Synthesized troubleshooting guides and runbooks from multiple sources create more efficient debugging processes.
從多個來源綜合整理的故障排除指南和操作手冊，能打造更有效率的除錯流程。

Top tips from the security engineering team
來自安全工程團隊的頂尖秘訣

Use custom slash commands extensively
廣泛使用自訂斜線指令

Security engineering uses 50% of all custom slash command implementations in the entire monorepo. These custom commands streamline specific workflows and speed up repeated tasks.
安全工程部門使用了整個單一程式庫中 50%的自訂斜線指令實作。這些自訂指令能簡化特定工作流程，並加速重複性任務。

Let Claude talk first  讓 Claude 先發言

Instead of asking targeted questions to generate 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.
他們不再針對特定問題要求生成程式碼片段，而是告訴 Claude Code「邊做邊提交」，讓它自主工作並定期檢查，從而產出更全面的解決方案。

Leverage it for documentation
善用其文件生成功能

Beyond coding, Claude Code excels at synthesizing documentation and creating structured outputs. The team provides writing samples and formatting preferences to get documents they can immediately use in Slack, Google Docs, and other tools to avoid interface switching fatigue.
除了編寫程式碼，Claude Code 在整合文件資料和創建結構化輸出方面表現出色。團隊提供寫作範例和格式偏好，就能直接取得適用於 Slack、Google Docs 等工具的現成文件，避免頻繁切換介面造成的疲勞。

Claude Code for inference
Claude Code 的推論應用

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.
推論團隊負責管理記憶系統，該系統在 Claude 讀取您的提示並生成回應時儲存資訊。團隊成員，尤其是機器學習領域的新手，可以廣泛運用 Claude Code 來彌補知識差距並加速工作進程。

Main Claude Code use cases
Claude Code 主要應用場景

Codebase comprehension and onboarding
程式碼庫理解與新手上路

The team relies heavily on Claude Code to quickly understand the architecture when joining a complex codebase. 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.
團隊在加入複雜程式碼庫時，高度依賴 Claude Code 來快速理解架構。與其手動搜尋 GitHub 儲存庫，他們會請 Claude 找出哪些檔案調用了特定功能，在幾秒內就能獲得結果，無需詢問同事或手動搜尋。

Unit test generation with edge case coverage
單元測試生成，涵蓋邊界案例

After writing core functionality, they ask Claude to write comprehensive unit tests. Claude automatically includes missed edge cases, completing what would normally take a significant amount of time and mental energy in minutes, acting like a coding assistant they can review.
在完成核心功能後，他們會請 Claude 撰寫全面的單元測試。Claude 會自動包含遺漏的邊界案例，將原本需要耗費大量時間和腦力的工作，在幾分鐘內完成，就像一個可供審查的編碼助手。

Machine learning concept explanation
機器學習概念解釋

Team members without a machine learning background 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%.
沒有機器學習背景的團隊成員依賴 Claude 來解釋模型特定的功能和設定。原本需要花費一小時進行 Google 搜尋和閱讀文件的工作，現在只需 10-20 分鐘，將研究時間減少了 80%。

Cross-language code translation
跨語言程式碼轉譯

When testing functionality in different programming languages, the team explains 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.
在不同程式語言中測試功能時，團隊只需說明想測試的內容，Claude 就能用所需語言（例如 Rust）寫出邏輯，省去為了測試目的而學習新語言的麻煩。

Command recall and Kubernetes management
指令回憶與 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.
無需記憶複雜的 Kubernetes 指令，他們直接詢問 Claude 正確語法，例如「如何取得所有 pod 或部署狀態」，就能獲得基礎設施工作所需的確切指令。

Team impact  團隊影響

Accelerated ML concept learning
加速機器學習概念理解

With Claude Code, their research time is reduced by 80%, and what historically took an hour of Google searching now takes 10-20 minutes.
有了 Claude Code，他們的研究時間減少了 80%，過去需要花一小時 Google 搜尋的內容，現在只需 10-20 分鐘。

Faster codebase navigation
更快速的程式碼庫導覽

The tool can help team members find relevant files and understand system architecture in seconds instead of relying on colleagues to share knowledge, often over the course of several days.
這個工具能幫助團隊成員在幾秒內找到相關檔案並理解系統架構，而不必依賴同事花費數天時間分享知識。

Comprehensive test coverage
全面的測試覆蓋率

Claude automatically generates unit tests with edge cases, relieving mental burden while maintaining code quality.
Claude 能自動生成包含邊界案例的單元測試，在減輕心智負擔的同時維持程式碼品質。

Language barrier elimination
消除語言隔閡

The team can implement functionality in unfamiliar languages like Rust without needing to learn it.
團隊無需學習就能用不熟悉的語言（如 Rust）實現功能。

Top tips from the Inference team
推論團隊的頂尖秘訣

Test knowledge base functionality first
先測試知識庫功能

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.
嘗試詢問各種問題，看看 Claude 是否能比 Google 搜尋更快回答。如果它更快且更準確，這將成為你工作流程中省時的寶貴工具。

Start with code generation
從程式碼生成開始

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.
給 Claude 具體指令並要求它編寫邏輯，然後驗證正確性。這有助於在處理更複雜任務前，先建立對工具能力的信任。

Use it for test writing
用於撰寫測試

Having Claude write unit tests relieves significant pressure from daily development work. Use this feature to maintain code quality without spending time thinking through all test cases manually.
讓 Claude 編寫單元測試能大幅減輕日常開發工作的壓力。使用這項功能可在不花時間手動構思所有測試案例的情況下，維持程式碼品質。

Claude Code for data science and ML engineering
Claude Code 用於資料科學與機器學習工程

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. Claude Code enables these teams to build production-quality analytics dashboards without becoming full-stack developers.
資料科學與機器學習工程團隊需要複雜的可視化工具來理解模型效能，但建構這些工具通常需要掌握不熟悉的語言和框架的專業知識。Claude Code 讓這些團隊無需成為全端開發者，就能打造符合生產品質的分析儀表板。

Main Claude Code use cases
Claude Code 主要應用場景

Building JavaScript/TypeScript dashboard apps
建構 JavaScript/TypeScript 儀表板應用程式

Despite knowing "very little JavaScript and TypeScript," the team uses Claude Code to build entire React applications for visualizing Reinforcement Learning (RL) model performance and training data. They give Claude control to write full applications from scratch, like a 5,000-line TypeScript app, without needing to understand the code themselves. 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.
儘管團隊成員「對 JavaScript 和 TypeScript 所知甚少」，他們仍運用 Claude Code 來建構完整的 React 應用程式，用於視覺化強化學習（RL）模型的效能與訓練數據。他們讓 Claude 全權負責從零開始編寫完整應用程式，例如一個 5,000 行的 TypeScript 應用程式，而無需自行理解程式碼。這項能力至關重要，因為視覺化應用程式相對而言所需背景知識較少，且無需理解整個單體式代碼庫，能快速原型化工具來理解模型在訓練與評估期間的表現。

Handling repetitive refactoring tasks
處理重複性的重構任務

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.
當遇到合併衝突或較複雜的文件重構（複雜到編輯器巨集無法處理，但又不足以投入大量開發資源時），他們會像玩「吃角子老虎機」一樣使用 Claude Code——提交當前狀態，讓 Claude 自主運作 30 分鐘，要麼接受解決方案，要麼在無效時重新開始。

Creating persistent analytics tools instead of throwaway notebooks
建立持久性分析工具而非一次性筆記本

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. 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."
團隊不再構建用完即棄的 Jupyter 筆記本，而是讓 Claude 建立可重複使用的 React 儀表板，供未來模型評估使用。這點至關重要，因為理解 Claude 的表現是「團隊最重要的工作之一」——他們需要了解模型在訓練和評估期間的表現，而「這實際上並非易事，簡單工具無法從單一數值變化中獲取太多訊號」。

Zero-dependency task delegation
零依賴任務委派

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. This allows productivity in areas outside their expertise instead of spending time learning new technologies.
對於完全不熟悉的程式碼庫或語言任務，他們會將整個實作工作委派給 Claude Code，利用其能從單一儲存庫收集上下文並執行任務的能力，無需親自參與實際編碼過程。這讓他們能在專業領域外維持生產力，而不必花時間學習新技術。

Team impact  團隊影響

Achieved 2-4x time savings
節省 2-4 倍時間

Routine refactoring tasks that were tedious but manageable manually are now completed much faster.
原本繁瑣但尚可手動處理的例行重構任務，現在能更快完成。

Built complex applications in unfamiliar languages
用不熟悉的語言建構複雜應用程式

Created 5,000-line TypeScript applications despite having minimal JavaScript/TypeScript experience.
儘管只有極少的 JavaScript/TypeScript 經驗，仍成功開發出 5,000 行程式碼的 TypeScript 應用程式。

Shifted from throwaway to persistent tools
從一次性工具轉向持久性工具

Instead of disposable Jupyter notebooks, now building reusable React dashboards for model analysis.
不再使用一次性 Jupyter 筆記本，現在改為建構可重複使用的 React 儀表板來進行模型分析。

Direct model improvement insights
直接獲取模型改進的洞察

Firsthand Claude Code experience informs development of better memory systems and UX improvements for future model iterations.
親身體驗 Claude Code 的經驗，為未來模型迭代提供了改進記憶體系統和使用者體驗的寶貴參考。

Enabled visualization-driven decision making
啟用視覺化驅動的決策制定

Better understanding of Claude's performance during training and evaluations through advanced data visualization tools.
透過先進的資料視覺化工具，更深入理解 Claude 在訓練與評估期間的表現

Top tips from Data Science and ML Engineering teams
來自資料科學與機器學習工程團隊的頂尖技巧

Treat it like a slot machine
把它當作吃角子老虎機來看待

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.
在讓 Claude 開始工作前先保存你的狀態，讓它運行 30 分鐘，然後要麼接受結果，要麼重新開始，而不是試圖糾正錯誤。重新開始的成功率通常比試圖修正 Claude 的錯誤更高。

Interrupt for simplicity when needed
必要時中斷以追求簡潔

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.
在監督過程中，不要猶豫停止 Claude 並詢問「你為什麼要這樣做？試試更簡單的方法。」模型預設傾向於更複雜的解決方案，但對要求簡化方法的回應良好。

Claude Code for product engineering
產品工程用的 Claude Code

The Product Engineering team works on features like PDF support, citations, and web search that bring additional knowledge into Claude's context window. 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. Claude Code improves this experience by serving as a guide that can help them understand system architecture, identify relevant files, and explain complex interactions.
產品工程團隊負責開發如 PDF 支援、引用功能和網路搜尋等特性，這些功能能為 Claude 的上下文視窗帶來額外知識。在龐大且複雜的程式碼庫中工作，意味著經常會遇到不熟悉的程式碼段落，需要花費大量時間理解針對特定任務該檢視哪些檔案，並在進行修改前建立上下文理解。Claude Code 透過擔任引導者的角色改善了這個體驗，幫助團隊理解系統架構、識別相關檔案，並解釋複雜的互動關係。

Main Claude Code use cases
Claude Code 主要應用場景

First-step workflow planning
第一步工作流程規劃

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. This replaces the traditional time-consuming process of manually navigating the codebase and gathering context before starting work.
團隊將 Claude Code 作為任何任務的「第一站」，要求它識別針對錯誤修正、功能開發或分析需要檢視哪些檔案。這取代了傳統上耗時的手動瀏覽程式碼庫並在開始工作前收集上下文的過程。

Independent debugging across codebases
跨程式碼庫的獨立除錯

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? This is the behavior I'm seeing" and often get immediate progress, which wasn't feasible before given the time investment required.
團隊現在有信心能自行解決程式碼庫中不熟悉部分的錯誤，而不必求助他人。他們可以直接詢問 Claude「你覺得你能修復這個錯誤嗎？我觀察到的現象是這樣」，通常能立即獲得進展，這在過去需要投入大量時間的情況下是難以實現的。

Model iteration testing through dogfooding
透過內部使用進行模型迭代測試

Claude Code automatically uses the latest research model snapshots, making it their primary way of experiencing model changes. This gives the team direct feedback on model behavior changes during development cycles, which they hadn't experienced during previous launches.
Claude Code 會自動採用最新的研究模型快照，成為團隊體驗模型變化的主要途徑。這讓團隊在開發週期中能直接獲得模型行為變化的反饋，這是過去產品發布時未曾有過的體驗。

Eliminating context-switching overhead
消除上下文切換的開銷

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.
與其複製程式碼片段並將檔案拖曳至 Claude.ai，同時還要花費大量時間解釋問題，他們現在可以直接在 Claude Code 中提問，無需額外收集上下文資訊，大幅降低了心智負擔。

Team impact  團隊影響

Increased confidence in tackling unfamiliar areas
提升處理陌生領域的信心

Team members can independently debug bugs and investigate incidents in unfamiliar codebases.
團隊成員能夠獨立針對不熟悉的程式碼庫進行除錯與事件調查。

Significant time savings in context gathering
大幅節省收集上下文資訊的時間

Claude Code eliminated the overhead of copying code snippets and dragging files into Claude.ai, reducing mental context-switching burden.
Claude Code 消除了複製程式碼片段和將檔案拖曳至 Claude.ai 的繁瑣步驟，減輕了思維切換的負擔。

Faster rotation onboarding
加速輪調適應期

Engineers rotating to new teams can quickly navigate unfamiliar codebases and contribute meaningfully without extensive colleague consultation.
輪調至新團隊的工程師能快速熟悉陌生程式碼庫，無需頻繁諮詢同事即可做出實質貢獻。

Enhanced developer happiness
提升開發者幸福感

The team reports feeling happier and more productive with reduced friction in their daily workflows.
團隊回報表示，在日常工作流程中減少了摩擦，讓他們感到更快樂且更有生產力。

Top tips from the Product Engineering team
產品工程團隊的頂尖建議

Treat it as an iterative partner, not a one-shot solution
將其視為迭代合作的夥伴，而非一次性解決方案

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.
與其期待 Claude 能立即解決問題，不如將其視為一個可以反覆合作的夥伴。這種方式比第一次就試圖獲得完美解決方案更有效。

Use it for building confidence in unfamiliar areas
用於在不熟悉的領域建立信心

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.
別猶豫去解決那些超出你專業範圍的錯誤或調查事件。Claude Code 讓你能夠獨立處理通常需要大量背景知識建構的領域。

Start with minimal information
從最少的資訊開始

Begin with just the bare minimum of what you need and let Claude guide you through the process, rather than front-loading extensive explanations.
只需從最基本的資訊著手，讓 Claude 引導你完成整個流程，而不是一開始就塞入大量解釋。

Claude Code for growth marketing
成長行銷專用的 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. As a non-technical team of one, they use Claude Code to automate repetitive marketing tasks and create agentic workflows that would traditionally require significant engineering resources.
成長行銷團隊專注於建立跨付費搜尋、付費社群媒體、行動應用商店、電子郵件行銷和 SEO 的績效行銷管道。作為一人組成的非技術團隊，他們使用 Claude Code 來自動化重複性行銷任務，並創建傳統上需要大量工程資源的代理工作流程。

Main Claude Code use cases
Claude Code 主要應用場景

Automated Google Ads creative generation
自動化生成 Google Ads 廣告素材

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). 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. This has enabled them to test and iterate at scale, something that would have taken a significant amount of time to achieve previously.
團隊建立了一個自主工作流程，能處理包含數百則現有廣告與成效指標的 CSV 檔案，找出表現不佳的廣告進行迭代，並生成符合嚴格字數限制的新版本（標題限 30 字元，描述限 90 字元）。透過兩個專責子代理程式（一個負責標題，一個負責描述），系統能在數分鐘內產出數百則新廣告，無需手動跨多個廣告活動進行創作。這讓他們能大規模測試與迭代，過去要達成相同效果需耗費大量時間。

Figma plugin for mass creative production
Figma 外掛程式批量生產廣告素材

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 copy-pasting to half a second per batch. This enables 10x creative output, allowing the team to test vastly more creative variations across key social channels.
團隊開發了 Figma 外掛程式，取代手動複製編輯付費社群廣告靜態圖像的流程。該外掛能識別畫框，並透過替換標題與描述，以程式化方式批量生成最多 100 種廣告變體，將原本需要數小時的複製貼上作業，縮減至每批次僅需半秒鐘。這使創意產出提升 10 倍，讓團隊能在主要社群管道測試更多元化的廣告版本。

Meta Ads MCP server for campaign analytics
Meta Ads MCP 伺服器（用於廣告活動分析）

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.
他們建立了一個與 Meta Ads API 整合的 MCP 伺服器，可直接在 Claude 桌面應用程式中查詢廣告活動成效、花費數據和廣告效果，省去了在不同平台間切換進行成效分析的時間，這種效率提升能直接轉化為更好的投資回報率。

Advanced prompt engineering with memory systems
進階提示工程與記憶系統

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 self-improving testing framework. This enables systematic experimentation that would be impossible to track manually.
他們實作了一套基礎記憶系統，能記錄廣告迭代過程中的假設與實驗數據，讓系統在生成新變體時能參考過往測試結果，形成自我優化的測試框架。這使得系統化實驗成為可能，而這些實驗若以人工方式追蹤根本無法實現。

Team impact  團隊影響

Dramatic time savings on repetitive tasks
重複性任務的時間大幅縮減

Claude Code reduced ad copy creation time from 2 hours to 15 minutes, freeing up the team for more strategic work.
Claude Code 將廣告文案創作時間從 2 小時縮短至 15 分鐘，讓團隊能投入更具策略性的工作。

10x increase in creative output
創意產出量提升 10 倍

The team can now test vastly more ad variations across channels with automated ad generation and a Figma integration to source up-to-date visual design elements.
團隊現在可以透過自動化廣告生成和 Figma 整合來取得最新的視覺設計元素，跨管道測試更多廣告變體。

Operating like a larger team
像更大團隊一樣運作

The team can handle large development tasks that traditionally required dedicated engineering resources.
團隊能夠處理傳統上需要專門工程資源的大型開發任務。

Strategic focus shift  策略性焦點轉移

The team can spend more time on overall strategy and building agentic automation rather than manual execution.
團隊可以將更多時間投入整體策略和建構自主自動化系統，而非手動執行。

Top tips from the Growth Marketing team
成長行銷團隊的頂尖秘訣

Identify API-enabled repetitive tasks
識別支援 API 的重複性任務

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.
尋找那些涉及重複性操作且具備 API 的工作流程（例如廣告平台、設計工具、分析平台）。這些是最適合自動化的場景，也是 Claude Code 能發揮最大價值的地方。

Break complex workflows into specialized sub-agents
將複雜工作流程拆解為專業子代理

Instead of trying to handle everything in one prompt or workflow, create separate agents for specific tasks (like a headline agent vs. a description agent). This makes debugging easier and improves output quality when dealing with complex requirements.
不要試圖在單一提示或工作流程中處理所有事情，應為特定任務創建獨立代理（例如標題代理與描述代理分開）。這種方式能簡化除錯過程，並在處理複雜需求時提升輸出品質。

Thoroughly brainstorm and prompt plan before coding
在編碼前徹底進行腦力激盪與提示規劃

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.
預先花費大量時間使用 Claude.ai 來思考整個工作流程，然後讓 Claude.ai 為 Claude Code 創建一個全面的提示和代碼結構以供參考。此外，逐步進行工作，而不是要求一次性解決方案，以避免 Claude 因複雜任務而不堪重負。

Claude Code for product design
產品設計的 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 iteration with engineers.
產品設計團隊支援 Claude Code、Claude.ai 和 Anthropic API，專注於構建 AI 產品。即使是非開發人員也可以使用 Claude Code 來彌補傳統設計與工程之間的差距，使他們能夠直接實現設計願景，而無需與工程師進行大量的來回迭代。

Main Claude Code use cases
Claude Code 主要應用場景

Front-end polish and state management changes
前端優化與狀態管理變更

Instead of creating extensive design documentation and going through multiple rounds of feedback with engineers for visual tweaks (typefaces, colors, spacing), the team directly implements these changes using Claude Code. 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.
團隊不再需要撰寫繁瑣的設計文件，也不必與工程師反覆討論視覺調整（如字型、色彩、間距），而是直接運用 Claude Code 實作這些變更。工程師們發現，他們正在進行「通常不會由設計師操刀的大型狀態管理變更」，這讓他們能精準實現心目中的理想品質。

GitHub Actions automated ticketing
GitHub Actions 自動化票務系統

Using Claude Code's GitHub integration, they 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.
透過 Claude Code 的 GitHub 整合功能，他們只需提交描述所需變更的問題/工單，Claude 便會自動提出程式碼解決方案，無需開啟 Claude Code 介面，為持續累積的優化任務打造無縫的錯誤修復與功能精進工作流程。

Rapid interactive prototyping
快速互動式原型設計

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.
將設計草圖貼入 Claude Code 後，系統會生成完全可運作的原型，工程師能立即理解並進行迭代，取代傳統需大量解釋說明才能將靜態 Figma 設計轉化為實作程式碼的開發循環。

Edge case discovery and system architecture understanding
邊界案例探索與系統架構理解

The team uses 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.
團隊使用 Claude Code 來繪製錯誤狀態、邏輯流程和不同系統狀態，讓他們能在設計階段就識別出邊緣案例，而非等到開發後期才發現，從根本上提升了初始設計的品質。

Complex copy changes and legal compliance
複雜文案修改與法規遵循

For tasks like removing "research preview" messaging across the entire codebase, they use 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-and-forth coordination.
對於像是移除整個程式碼庫中「研究預覽」訊息這類任務，他們運用 Claude Code 找出所有相關實例、審查周邊文案、與法務部門即時協調變更並實施更新，這個過程僅需兩次 30 分鐘的電話會議即可完成，省去了原本需要一週來回協調的時間。

Team impact  團隊影響

Transformed core workflow
核心工作流程轉型

Claude Code became a primary design tool, with Figma and Claude Code open 80% of the time.
Claude Code 成為主要設計工具，Figma 和 Claude Code 有 80%的時間都開著。

2-3x faster execution  執行速度快 2-3 倍

Visual and state management changes that previously required extensive back-and-forth with engineers are now implemented directly.
過去需要與工程師反覆溝通的視覺和狀態管理變更，現在可以直接實作。

Weeks to hours cycle time
週期從數週縮短至數小時

Complex projects like Google Analytics launch messaging that would take a week of coordination are now completed in two 30-minute calls.
像 Google Analytics 發布訊息這類複雜專案，原本需要一週協調時間，現在只需兩次 30 分鐘的會議就能完成。

Two distinct user experiences
兩種截然不同的使用者體驗

Developers get an "augmented workflow" (faster execution), while non-technical users get a "holy crap, I'm a developer workflow.”
開發者獲得「增強型工作流程」（執行速度更快），而非技術使用者則體驗到「天啊，我竟然也能像開發者一樣工作」的流程。

Improved design-engineering collaboration
提升設計與工程團隊的協作效率

Claude Code enables improved communication and faster problem-solving because designers understand system constraints and possibilities without having to work as closely with engineers.
Claude Code 促進了更好的溝通與更快速的問題解決，因為設計師無需與工程師密切合作，就能理解系統的限制與可能性。

Top tips from the product design team
產品設計團隊的頂尖秘訣

Get proper setup help from engineers
從工程師那裡獲得適當的設定協助

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.
請工程團隊成員協助初始儲存庫設定與權限管理——技術導入對非開發人員來說具有挑戰性，但一旦配置完成，將為日常工作流程帶來變革性的影響。

Use custom memory files to guide Claude's behavior
使用自訂記憶檔案來引導 Claude 的行為

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.
建立具體指令告訴 Claude 你是一位缺乏編碼經驗的設計師，需要詳細解釋與較小規模的漸進式修改，這能顯著提升 Claude 回應的品質並降低使用門檻。

Leverage image pasting for prototyping
利用圖片貼上功能進行原型設計

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 直接將截圖貼到 Claude Code 中。它擅長解讀設計並生成功能性程式碼，對於將靜態模型轉化為工程師能立即理解並進一步開發的互動式原型來說，這項功能極具價值。

Claude Code for RL Engineering
Claude Code 在強化學習工程中的應用

The RL Engineering team focuses on efficient sampling in RL and weight transfers across the cluster. 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.
強化學習工程團隊專注於 RL 中的高效採樣和跨叢集的權重轉移。他們主要使用 Claude Code 來開發中小型功能、除錯以及理解複雜的程式碼庫，採用包含頻繁檢查點和回滾的迭代式開發方法。

Main Claude Code use cases
Claude Code 主要應用場景

Feature development with supervised autonomy
具監督自主性的功能開發

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. They work interactively, allowing Claude to take the lead but steering it when it goes off track.
團隊讓 Claude Code 負責撰寫中小型功能的大部分程式碼，同時進行監督，例如為重量轉移元件實作驗證機制。他們採用互動式工作模式，讓 Claude 主導開發流程，但在偏離方向時適時導正。

Test generation and code review
測試生成與程式碼審查

After implementing changes themselves, the team asks Claude Code to add tests or review their code. This automated testing workflow saves significant time on routine but important quality assurance tasks.
團隊在自行實作變更後，會要求 Claude Code 添加測試或審查程式碼。這種自動化測試工作流程為例行但重要的品質保證任務節省了大量時間。

Debugging and error investigation
除錯與錯誤調查

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.
他們使用 Claude Code 來除錯錯誤，效果有好有壞。有時它能立即識別問題並添加相關測試，但有時則難以理解問題所在，不過整體而言在運作順暢時仍能提供價值。

Codebase comprehension and call stack analysis
程式碼庫理解與呼叫堆疊分析

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.
他們工作流程中最大的改變之一，就是使用 Claude Code 快速獲取相關元件與呼叫堆疊的摘要，取代了手動閱讀程式碼或生成大量除錯輸出的方式。

Kubernetes operations guidance
Kubernetes 操作指引

They frequently ask Claude Code about Kubernetes operations that would otherwise require extensive Googling or asking their counterparts in Infrastructure Engineering, getting immediate answers for configuration and deployment questions.
他們經常向 Claude Code 詢問關於 Kubernetes 操作的相關問題，這些問題原本需要大量 Google 搜尋或詢問基礎設施工程團隊的同事，現在能立即獲得配置與部署問題的解答。

Development workflow impact
開發工作流程影響

Experimental approach enabled
實驗性方法已啟用

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.
他們現在採用「嘗試與回滾」的方法論，頻繁地提交檢查點，這樣就能測試 Claude 自主實作的嘗試，並在需要時回退，讓開發更具實驗性。

Documentation acceleration
文件加速

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 會自動添加實用的註解，大幅節省文件編寫時間，不過他們也注意到它偶爾會在奇怪的位置添加註解，或使用有疑慮的程式碼組織方式。

Speed-up with limitations
有限制的加速

While Claude Code can implement small-to-medium PRs with "relatively little time" from them, they acknowledge it only works on the first attempt about one-third of the time, requiring either additional guidance or manual intervention.
雖然 Claude Code 能在「相對較短時間」內完成中小型 PR（Pull Request）的實作，但團隊承認僅約三分之一的情況能在首次嘗試就成功，需要額外指導或手動介入。

Top tips from the RL Engineering team
RL 工程團隊的頂級建議

Customize your Claude.md file for specific patterns
為特定模式自訂你的 Claude.md 檔案

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 output consistency.
在你的 Claude.md 檔案中加入指令，防止 Claude 重複犯下工具呼叫的錯誤，例如告訴它「直接執行 pytest 而非先執行 run，且不要無謂地切換目錄 - 只需使用正確路徑即可」。這能顯著提升輸出的一致性。

Use a checkpoint-heavy workflow
採用檢查點密集的工作流程

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.
隨著 Claude 進行修改，定期提交你的工作成果，這樣當實驗不如預期時就能輕鬆回退。這種做法讓開發過程能更勇於嘗試，同時避免風險。

Try one-shot first, then collaborate
先嘗試單次示範，再進行協作

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.
先給 Claude 一個簡短的提示，讓它嘗試完整實現。如果成功（大約有三分之一的機率），你就節省了大量時間。如果不行，再轉向更協作、有引導的方式。

Claude Code for Legal  Claude Code 在法律領域的應用

The Legal team discovered Claude Code's potential through experimentation and a desire to learn about Anthropic's product offerings. 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.
法律團隊透過實驗和了解 Anthropic 產品線的渴望，發現了 Claude Code 的潛力。此外，團隊中一名成員有個個人使用案例，涉及為家庭和工作原型創建無障礙工具，這些原型展示了這項技術對非開發者的強大功能。

Main Claude Code use cases
Claude Code 主要應用場景

Custom accessibility solution for family members
為家庭成員打造的客製化無障礙解決方案

Team members have built communication assistants for family members with speaking difficulties due to medical diagnoses. In just one hour, one individual 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.
團隊成員為因醫療診斷而有語言障礙的家庭成員開發了溝通輔助工具。僅花費一小時，一位成員就利用原生語音轉文字技術打造出預測文字應用程式，能建議回覆內容並透過語音庫發聲，彌補了語言治療師推薦的現有無障礙工具不足之處。

Legal department workflow automation
法律部門工作流程自動化

The team 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.
團隊建立了原型「電話樹」系統，協助成員聯繫 Anthropic 內部合適的律師，展現法律部門如何在不依賴傳統開發資源的情況下，為常見任務打造客製化工具。

Team coordination tools  團隊協作工具

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.
管理者們建立了 G Suite 應用程式，能自動化每週團隊更新並追蹤各產品的法律審查狀態，讓律師們只需簡單點擊按鈕就能快速標記需要審查的項目，而無需管理繁瑣的試算表。

Rapid prototyping for solution validation
快速原型製作以驗證解決方案

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.
他們運用 Claude Code 快速建構功能性原型，可向領域專家展示（例如向 UCSF 專家展示無障礙工具），在投入更多時間前驗證構想並找出現有解決方案。

Work style and impact  工作方式與影響力

Planning in Claude.ai, building in Claude Code
在 Claude.ai 上規劃，於 Claude Code 中建構

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.
他們採用兩階段流程：先透過 Claude.ai 進行腦力激盪與規劃，接著轉移到 Claude Code 進行實作，過程中會要求系統放慢速度逐步作業，而非一次輸出所有內容。

Visual-first approach  視覺優先的作法

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.
他們經常使用截圖向 Claude Code 展示想要的介面樣貌，然後根據視覺回饋進行迭代，而非用文字描述功能。

Prototype-driven innovation
原型驅動的創新

They emphasize overcoming the fear of sharing "silly" or "toy" prototypes, as these demonstrations inspire others to see possibilities they hadn't considered.
他們強調要克服分享「幼稚」或「玩具般」原型的恐懼，因為這些示範能激發他人看見未曾考慮的可能性。

Security and compliance awareness
安全與合規意識

MCP integration concerns  MCP 整合疑慮

Product lawyers use Claude Code to immediately identify security implications of deep MCP integrations, noting how conservative security postures will create barriers as AI tools access more sensitive systems.
產品法律團隊運用 Claude Code 即時識別深度 MCP 整合的安全影響，並指出當 AI 工具存取更敏感的系統時，保守的安全態勢將形成障礙。

Compliance tooling priorities
合規工具優先事項

They advocate for building compliance tools quickly as AI capabilities expand, recognizing the balance between innovation and risk management.
他們主張隨著 AI 能力的擴展應快速建立合規工具，同時認知到創新與風險管理之間的平衡。

Top tips from the Legal Department
來自法律部門的頂尖建議

Plan extensively in Claude.ai first
先在 Claude.ai 上進行全面規劃

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.
在轉移到 Claude Code 之前，先使用 Claude 的對話介面來完善你的整個構想。然後請 Claude 將所有內容總結成逐步實施的提示。

Work incrementally and visually
逐步且視覺化地工作

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.
請 Claude Code 放慢速度，一次只實作一個步驟，這樣你就能輕鬆複製貼上而不會感到不知所措。盡量使用截圖來展示你希望介面呈現的樣子。

Share prototypes despite imperfection
儘管不夠完美，仍分享原型

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.
克服隱藏「玩具」專案或未完成作品的衝動。分享原型能讓其他人看見可能性，並激發平常少有互動的部門之間的創新火花。

Get started with Claude Code today.
立即開始使用 Claude Code。