How Anthropic Dogfoods On Claude Code 

Explain, Refactor, Generate

Claude Code’s central functions include its ability to explain code, to refactor code (to correct errors, to improve maintainability, to standardize code style across distributed teams to achieve more harmonious collaboration, or to eliminate wasteful coding practices and so reduce technical debt) and also to generate new code. The software itself embeds Claude Opus 4, Anthropic’s AI model for building what it calls “frontier intelligence”, which has “deep codebase awareness” and the ability to edit files and run commands directly in a developer’s own environment. 

Claude Opus 4 is the model Anthropic’s own researchers and engineers use, so there’s the dogfooding factor right there. For external users, it’s available in Claude for Pro, Max, Team and Enterprise user licences. Claude Code is suitable for non-technical staff looking to automate tasks and untangle project complexity. 

Anthropic says that internally, it sees Claude Code being used by data infrastructure teams, product development and security engineers and those working on inference logic. It also sees the technology in the hands of its growth marketing division, product design specialists, reinforcement learning engineering and legal.

Kubernetes Debugging With Screenshots

When Anthropic’s own 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,” noted the Anthropic blog.

Starting, logically perhaps, closer to the software engineering team, Anthropic has also detailed work with Claude Code outside the realm of its technical team; accounting staff have been using plain text files to describe their data workflows and get fully automated execution. Non-coders can issue commands such as “query this dashboard, get information, run these queries, produce Excel output,” and Claude Code executes the entire workflow. It also asks for any additional required inputs, like dates.

Claude.md: An Opinionated Control Panel

When new AI and data scientists join up at Anthropic, they are asked to use Claude Code to navigate the massive codebase in front of them. Claude Code reads their Claude.md files (documentation), identifies relevant files for specific tasks, explains data pipeline dependencies and works to help rookies (experienced, or relative novices) to understand which upstream sources feed into dashboards. The team says that this replaces traditional data catalogs and discoverability tools.

According to independent coder Anthony Calzadilla, front-end web manager at Certinia, “If you’re building with Claude Code, the claude.md file is your control panel. It’s how you set constraints, establish the project structure, and teach the AI how to operate within your stack, without bloating the codebase or relying on fragile comments. Think of it as a highly opinionated prompt that rides along with every request, telling Claude what to do, what not to touch… and how to keep the output production-ready.”

The Anthropic team asks Claude Code to summarize completed work sessions and suggest improvements at the end of each task it carries out. This, it says, creates a “continuous improvement loop” where Claude Code helps refine the Claude.md documentation at hand. It also improves the workflow instructions being executed based on actual usage, making subsequent iterations more effective.

In terms of practical usefulness, Claude Code can process much larger data volumes and identify anomalies (such as monitoring 200 dashboards, for example) that would be impossible for humans to review manually.

How to Use Claude Code

According to the team, the better a developer can document 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 a software engineer has existing design patterns. Anthropic also recommends using MCP servers rather than the BigQuery CLI to achieve and maintain better security control over what Claude Code can access. This is especially the case when handling sensitive data that requires logging or that has potential privacy concerns.

“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. After writing the 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,” blogged Anthropic.

How to Use Cloud Code, Really

For some practical, hands-on, real-world use case advice, Anthropic says Claude Code should be regarded as the “first stop” for any task, asking it to identify which files to examine for bug fixes, feature development, or analysis. This technology uses the latest research model snapshots, making it a primary means of experiencing and consuming model changes. This gives the team direct feedback on model behavior changes during development cycles, which they hadn’t experienced during previous launches.

The end result (at Anthropic at least) is happier coders and happier business practitioners. Not to temper that happiness factor too much, Anthropic urges users to be realistic and treat Claude Code as an “iterative partner” and not as a one-shot solution or total panacea of any kind. The advice is to begin with just the bare minimum and let Claude Code (and indeed Claude) guide users through a process, rather than front-loading extensive explanations. If users approach this service as collaborators with whom they iterate, they will do better than if they expect it to get perfect solutions on the first try.