How Anthropic Uses Claude Code Internally

When a company builds a developer tool, the most revealing question you can ask is: do they actually use it themselves? In Anthropic's case, the answer is an emphatic yes. Claude Code is not just a product Anthropic ships to customers -- it is a tool that virtually every team inside the company relies on daily, from the engineers building Claude's core models to the marketing team generating advertising campaigns and the legal department prototyping internal applications.

What makes Anthropic's internal usage particularly instructive is the breadth of teams involved. This is not a story about software engineers using a coding tool for software engineering. It is a story about how an agentic AI development tool transforms workflows across an entire organization, including for people who have never written a line of production code in their lives.

The data comes directly from Anthropic's own published account of how their teams use Claude Code. The numbers are real. The workflows are battle-tested in production. And the lessons are directly applicable to any team considering adopting Claude Code or scaling their existing usage.

Let us walk through each department, examine their specific workflows, and extract the patterns that make Claude Code so effective across such different contexts.

Codebase Navigation and Onboarding

Every developer has experienced the disorientation of joining a new team or project. You open the repository, stare at hundreds of files organized in ways that made sense to someone else, and spend hours trying to understand which parts of the code are relevant to your task. At Anthropic, this problem is compounded by the complexity of their AI infrastructure -- the codebase spans model training pipelines, inference systems, API services, safety tooling, and internal applications.

Anthropic's Product Engineering team has turned Claude Code into their first stop for understanding unfamiliar parts of the codebase. Rather than spending 30 minutes tracing import chains and reading through file after file, engineers ask Claude Code to explain how a particular feature works, which files are involved, and how data flows through the system. Claude Code reads the relevant source files, understands the relationships between components, and provides a coherent explanation in seconds.

The Infrastructure team's data scientists take this a step further by maintaining comprehensive CLAUDE.mdfiles in their repositories. These files serve as structured onboarding documents that Claude Code reads automatically when you enter a project directory. They contain information about project conventions, architecture decisions, environment setup, common tasks, and team-specific patterns. When a new data scientist joins the team and opens the repository, Claude Code already understands the context and can answer questions like "how do I run the model evaluation pipeline?" or "where are the configuration files for the data processing jobs?" without anyone needing to schedule an onboarding session.

This pattern -- using CLAUDE.md as a living onboarding document -- has become one of the most widely replicated practices across the Claude Code community. It costs almost nothing to maintain (you update it as part of your normal development workflow) and dramatically reduces the time new team members need to become productive.

Testing and Code Review

Two teams at Anthropic have developed particularly sophisticated testing workflows with Claude Code: Product Design and Security Engineering.

Automated PR Comments via GitHub Actions

The Product Design team built a GitHub Actions integration where Claude Code automatically reviews every pull request and posts detailed comments. This is not a simple linting check -- Claude Code reads the full diff, understands the intent of the changes in context of the broader codebase, and provides substantive feedback on code quality, potential bugs, missing edge cases, and architectural concerns. The comments appear directly on the PR, making code review faster and more thorough without requiring a human reviewer to do the initial pass.

This workflow is particularly valuable for teams with limited senior reviewer bandwidth. Junior engineers get immediate, high-quality feedback on their changes while senior engineers can focus their review time on the architectural and strategic aspects that require human judgment.

Test-Driven Development Transformation

The Security Engineering team took a different approach. They transformed their entire development workflow to embrace test-driven development (TDD) with Claude Code as the implementation engine. The workflow looks like this: a security engineer writes a failing test that defines the expected behavior, then hands the test to Claude Code and asks it to write the implementation that makes the test pass.

This is a fundamentally different way of using an AI coding tool. Instead of asking Claude Code to write code from a natural language description (which can be ambiguous), the engineer expresses their intent in the most precise language available -- executable test cases. Claude Code then has an unambiguous success criterion: make the tests pass. The result is code that is correct by construction, well-tested from the start, and aligned with the engineer's exact intent.

The Security team reports that this workflow produces higher quality code with fewer bugs, catches regressions earlier, and has significantly improved their test coverage across the security infrastructure.

Production Debugging

Production incidents are high-pressure situations where speed matters. Every minute of downtime costs money, erodes user trust, and stresses the team. Two of Anthropic's most technically demanding teams have found that Claude Code dramatically accelerates their incident response.

Security Engineering: 3x Faster Stack Trace Resolution

When a production error occurs, the Security Engineering team feeds the stack trace directly to Claude Code. Claude Code reads the stack trace, locates the relevant source files in the codebase, traces the execution path that led to the error, identifies the root cause, and often suggests a fix -- all in a single interaction. What previously required an engineer to manually navigate through multiple files, read documentation, check recent changes, and piece together the causal chain now happens in a fraction of the time.

The team reports that this workflow resolves production issues approximately 3x faster than their previous manual debugging process. In the context of a production incident, this means the difference between a 30-minute outage and a 10-minute outage -- a meaningful improvement for both the team and their users.

Data Infrastructure: Diagnosing Kubernetes Issues from Screenshots

Perhaps the most impressive debugging story comes from the Data Infrastructure team. During a production outage, an engineer took a screenshot of a Kubernetes monitoring dashboard showing anomalous pod behavior and pasted it directly into Claude Code. Claude Code analyzed the visual information in the dashboard, identified the pattern as Kubernetes pod IP exhaustion, explained the root cause, and recommended specific remediation steps.

This single interaction saved the team approximately 20 minutesduring an active outage -- time that would have been spent manually querying Kubernetes APIs, checking IP allocation tables, and correlating logs. The ability to go from a dashboard screenshot to a diagnosed root cause in seconds demonstrates how Claude Code's multimodal capabilities (understanding both code and images) create workflows that were simply not possible with text-only tools.

Feature Development and Prototyping

Building new features is where Claude Code's agentic capabilities truly shine. Two use cases from Anthropic illustrate how the tool bridges the gap between design intent and working code, and between domain expertise and implementation capability.

From Figma to Code

The Product Design team feeds Figma design files directly to Claude Code and asks it to implement the designs as working frontend components. Claude Code analyzes the visual layout, typography, spacing, colors, and interactive elements in the design, then generates production-quality code that matches the design intent. This workflow eliminates the traditional back-and-forth between designers and frontend developers over pixel-level details and significantly accelerates the design-to-implementation pipeline.

Designers can iterate on their Figma files and re-generate the implementation in minutes rather than waiting for a developer to pick up the ticket, understand the design, implement it, request design review, and iterate. The feedback loop shrinks from days to minutes.

Data Scientists Building Full React Applications

One of the most striking examples of Claude Code's democratizing effect comes from Anthropic's data scientists. Team members who had never written TypeScript built full React applications for visualizing model performance metrics. These are not simple scripts or notebooks -- they are interactive web applications with charts, filters, real-time data updates, and polished user interfaces.

Previously, a data scientist who wanted a custom visualization tool had two options: build a basic Jupyter notebook (limited interactivity, hard to share) or file a ticket with the engineering team and wait weeks for it to be prioritized. With Claude Code, they describe what they want in terms of their domain expertise ("I need a dashboard that shows model accuracy over time, broken down by evaluation category, with the ability to compare different model versions") and Claude Code handles the TypeScript, React, charting library integration, and state management they do not know. The result is that domain experts can build the tools they need without learning an entirely new technology stack.

Documentation and Research

The Inference team at Anthropic works on some of the most technically complex systems in the company -- the infrastructure that serves Claude's responses to millions of users. Their work frequently requires deep research into specific technical topics: understanding a particular GPU optimization technique, evaluating a new quantization method, or investigating how a competitor's inference stack achieves certain performance characteristics.

Before Claude Code, this research process typically took about one hour per topic. An engineer would read academic papers, scan blog posts, search through internal documentation, and synthesize the information into a format useful for making engineering decisions.

With Claude Code, the same research process takes 10 to 20 minutes -- an 80% reduction. Engineers describe their research question to Claude Code, which draws on its training knowledge to provide a structured overview of the topic, summarize relevant approaches, identify key trade-offs, and point to specific implementation considerations. The engineer then uses this as a starting point for deeper investigation rather than starting from a blank page.

This pattern is not about replacing rigorous research. It is about accelerating the initial phase of gathering context and understanding the landscape, so the engineer can spend their time on the high-value work of evaluating, deciding, and implementing rather than the low-value work of finding and organizing information.

Marketing and Automation

Perhaps the most surprising Claude Code power user at Anthropic is the Growth Marketing team. This is a team of marketers, not engineers, and they have built one of the most sophisticated Claude Code workflows in the entire company.

The team created an agentic system that takes CSV files containing campaign parameters -- target audiences, value propositions, tone guidelines, channel specifications -- and generates hundreds of advertising variations automatically. Each ad variation is tailored to the specific audience segment, channel format, and messaging strategy defined in the input data.

Before this system, creating advertising variations was a manual process. A copywriter would write a handful of variations, the team would review them, request changes, and iterate. Producing 20 variations for a single campaign might take an entire day. The agentic system generates hundreds of variations in minutes, giving the team dramatically more creative options to test and optimize.

The key insight here is that the marketing team did not ask engineering to build this system for them. They built it themselves using Claude Code. The system involves reading CSV files, processing structured data, applying templates, generating text variations, and writing output files -- tasks that are fundamentally programming tasks, but expressed in terms the marketing team understands.

This is Claude Code's most underappreciated capability: it makes non-engineers productive with programming tasks by translating domain expertise into working code. The marketers did not need to learn Python or JavaScript. They needed to know what good advertising looks like, and Claude Code handled the rest.

Legal and Non-Engineering Use Cases

The Legal team's use of Claude Code is perhaps the clearest demonstration that this tool has transcended its origins as a "coding assistant." Anthropic's lawyers used Claude Code to build a prototype phone tree system that connects employees to the appropriate lawyer based on their question type.

Think about what this means. A team of lawyers, with no engineering background, used an AI coding tool to prototype a functional internal application. They described the workflow they wanted (employee calls in, answers a few routing questions, gets connected to the right legal specialist), and Claude Code generated the code to make it work.

This is not an isolated experiment. It reflects a broader pattern at Anthropic where non-engineering teams use Claude Code to build small internal tools, automate repetitive processes, and prototype ideas that would previously have required filing a ticket with engineering and waiting weeks for prioritization.

The organizational impact of this pattern is significant. When every team can build their own tools, the engineering team is freed from a backlog of internal tool requests and can focus on the core product. Non-engineering teams get exactly what they need, when they need it, without the translation losses that occur when a non-technical person describes a requirement to an engineer who interprets it through their own assumptions.

Productivity Gains Overview

Here is a consolidated view of the measurable productivity improvements Anthropic has documented across their teams:

These improvements are not theoretical. They come from daily usage by teams working on production systems at scale. The common thread is that Claude Code eliminates the low-value, time-consuming phases of each workflow (searching, reading, correlating, boilerplating) so that humans can focus on the high-value phases (deciding, designing, validating, and shipping).

The Key Lesson

The overarching lesson from Anthropic's internal adoption is not about any specific technical trick or configuration. It is about mindset. The teams that get the most value from Claude Code are not the ones that try to replace human work with AI work. They are the ones that identify the specific phases of their existing workflows where humans spend time on tasks that do not require human judgment, and then deploy Claude Code precisely there.

The Security team did not ask Claude Code to design their security architecture. They asked it to implement code that passes their carefully written tests. The Inference team did not ask Claude Code to make research decisions. They asked it to accelerate the information-gathering phase so they could make decisions faster. The Marketing team did not ask Claude Code to develop their brand strategy. They asked it to generate ad variations based on a strategy they had already defined.

In every case, the human remains the decision-maker, the quality controller, and the creative director. Claude Code handles the execution, the repetitive tasks, and the technical translation that previously created bottlenecks. This is what sustainable AI adoption looks like -- not replacement, but augmentation of human capabilities.

How to Apply These Patterns to Your Team

Based on Anthropic's experience, here are practical steps you can take to replicate their results:

• 1.Start with CLAUDE.md files. Create a comprehensive CLAUDE.md in every repository. Include project architecture, conventions, common tasks, and team-specific patterns. This single step dramatically improves Claude Code's effectiveness and accelerates onboarding for both humans and AI.
• 2.Adopt test-driven workflows. Write your tests first, then let Claude Code implement the code. This gives Claude Code an unambiguous success criterion and produces better code than natural language descriptions alone.
• 3.Integrate into CI/CD. Set up Claude Code as an automated PR reviewer via GitHub Actions. This catches issues early and reduces the burden on senior engineers to do initial code review passes.
• 4.Create incident response playbooks. Train your team to paste stack traces and dashboard screenshots into Claude Code during production incidents. The 3x speedup Anthropic achieved is directly replicable.
• 5.Empower non-engineering teams. Identify marketing, legal, operations, or data teams that have repetitive workflows or need internal tools. Claude Code can help them build their own solutions without engineering tickets.
• 6.Build agentic workflows for repetitive content. If your team generates variations of content (ads, emails, documentation, reports), build a Claude Code workflow that takes structured input and produces variations. The marketing team's CSV-to-ads pipeline is a pattern that generalizes to many content types.
• 7.Measure before and after. Track time spent on specific tasks before and after Claude Code adoption. Concrete metrics like "debugging time reduced from 30 minutes to 10 minutes" build organizational confidence and justify broader adoption.

Frequently Asked Questions

Related Resources

• --Claude Code Agents Directory-- Browse and discover pre-built agents for automating workflows.
• --Claude Code Skills Directory-- Find reusable skills to teach Claude Code your team's patterns.
• --The Complete Guide to Claude Code Agents-- Learn how to build multi-agent systems like Anthropic's marketing team.