Anthropic reveals how its internal team uses Claude Code (with full manual)

Original document:

https://www-cdn.anthropic.com/58284b19e702b49db9302d5b6f135ad8871e7658.pdf

Chinese translation:

https://k8274sonr5.feishu.cn/docx/GeIWdv2nCoVzxgxcCdlc7a0rnob?from=from_copylink

Practical usage by different teams

1. Technical and engineering teams (data infrastructure, product development, security, reasoning, RL engineering)

Such teams mainly use Claude Code to accelerate the existing software development life cycle. The main usage scenarios include:

Codebase understanding and onboarding : New employees or engineers who need to work in an unfamiliar codebase will ask Claude Code to explain the system architecture, identify key files and dependencies, and significantly shorten the onboarding time.

Function Development and Prototyping :

• Supervisory development : When developing core business logic, engineers will provide detailed instructions and supervise Claude Code in real time to ensure quality and compliance.
  

• Autonomous development : For non-core functions (such as implementing Vim mode) or rapid prototyping, engineers will enable "auto-acceptance mode" to let Claude Code write, test and iterate autonomously, and only do the final review and fine-tuning.

Automated testing : After completing feature development, engineers will ask Claude Code to write comprehensive unit tests and automatically cover boundary conditions that are easy to overlook.

Code Review and Refactoring : Used to review infrastructure code (such as Terraform plans) to assess risk, or to handle refactoring tasks that are too complex for editor macros but not complex enough to warrant significant development effort.

Debugging and incident response : When encountering complex problems (such as Kubernetes cluster failures), the team will provide Claude Code with stack traces, error logs, and even UI screenshots to guide it to find the root cause of the problem and provide repair instructions. In this way, the security engineering team has shortened the infrastructure debugging time from 10-15 minutes to about 5 minutes.

2. Data Science and Visualization Team

This team is quite special. Strictly speaking, it belongs neither to the previous technical and engineering teams nor to the non-technical team to be introduced below, so it is singled out.

They needed sophisticated visualization tools to understand model performance, but building such tools typically requires specialized skills in unfamiliar languages ​​and frameworks. Claude Code enabled this team to build production-quality analytical dashboards without having to become full-stack developers.

Build professional-grade applications : Even though team members knew “little” about JavaScript/TypeScript, they successfully used Claude Code to build a React application with thousands of lines of code for visual analysis of model performance.

From disposable to persistent : Instead of relying on disposable Jupyter notebooks, they create reusable, persistent React dashboards for ongoing evaluation of future models.

3. Non-technical teams (product design, growth marketing, legal)

The usage of this type of team reflects the core value of Claude Code, "empowerment", allowing them to independently complete tasks that in the past required the support of engineers.

Get started with product design : Designers no longer just deliver static models, but use Claude Code to fine-tune the front-end visuals (such as fonts, colors), and even make complex state management changes. They can also quickly generate interactive prototypes by pasting model diagrams.

Marketing Automation :

• Ad creative generation : The Growth Marketing team created an automated process that analyzes existing ad data and generates new ad copy at scale that meets strict character limits, reducing hours of work to minutes.

• Batch production of design materials : Developed a Figma plug-in to programmatically generate hundreds of advertising image variations with one click, increasing creative output by 10 times.

Prototyping Legal Tools : Legal team members built a customized assistive communication app for family members with speech disabilities and prototyped a “phone tree” system for the department to help colleagues quickly find the right attorney.

Differences and similarities in practical methods in different scenarios

1. Commonalities

1) Iterative collaboration : Almost all teams emphasize that Claude Code should not be seen as a one-time problem-solving tool, but a partner that requires iterative collaboration. Whether it is guiding it, correcting it, or reviewing it together, this collaborative model is the key to success.

2) Context is crucial : Providing high-quality context is a prerequisite for obtaining good output. This includes writing detailed Claude.md Documentation to explain workflows and tools, create clear and specific prompts, and use custom commands to simplify repetitive tasks.

3) Two-step workflow : Many teams, especially non-technical teams, adopt a two-step process of "planning first, then execution". They first fully conceive and plan the entire workflow in Claude.ai's conversational interface, then let it generate a comprehensive, step-by-step prompt, and finally hand it over to Claude Code to execute the coding.

4) Automating repetitive tasks : All teams use Claude Code to automate boring but important tasks, such as writing tests, batch generating content, and code formatting, thereby freeing up manpower to focus on higher-level strategic work.

Differences

1) Working mode: Enhancement vs. empowerment

Developers (Enhancement) : For engineers, Claude Code is an “enhanced workflow” that makes things that can already be done faster and more efficiently.

Non-developers (empowerment) : For non-technical personnel such as designers and legal affairs, it brings the experience of "Oh my God, I am a developer" and gives them new capabilities that they did not have before.

2) Degree of supervision: synchronous vs asynchronous

Concurrent supervision : When working on core business logic or critical fixes, the product development team conducts real-time and close supervision to ensure code quality.

Asynchronous autonomy : When working on prototyping, non-core features, or certain refactoring tasks, the team takes a more hands-off approach. For example, the product development team uses the "auto-accept mode" to let it run autonomously, while the data science team adopts the "slot machine" strategy - let it work autonomously for 30 minutes and either accept the results or start over.

3) Interaction method: code instructions vs visual instructions

The technical team prefers to provide specific code and logical instructions.

Non-technical teams (such as product design and legal) use visual input more frequently, such as pasting screenshots directly to show what they want the interface to look like, and then iterate through visual feedback. They also need Claude Code to slow down and explain step by step to make it easier to understand and follow.

4) Risk control strategy

The RL engineering team adopted a more cautious "checkpoint-intensive workflow" where they frequently submitted code so that they could easily roll back if Claude Code's attempt was unsuccessful. They also pointed out that the probability of success in one go was only about one-third, and they needed to be prepared for collaborative guidance. This reflects a stronger sense of risk management in scenarios with low fault tolerance.

Although this manual is used by Anthropic to show off its strengths, it is a good show. Their sharing of Claude Code's practice in different business scenarios actually shows what AI programming can do and what it cannot do.

Everyone should also be able to feel that AI programming is changing the traditional way of collaboration. People no longer need to wait for a perfect solution, but are more inclined to "quickly make a sample first, and then iterate and improve it step by step." This makes the entire collaboration method more flexible and bold, and can quickly turn various good ideas into valuable actual products.

Change is happening, whether we accept it or not.