The Data Infrastructure team at Anthropic does some of the most invisible but essential work in the company: organizing all business data for every team across the organization. They manage pipelines, dashboards, data quality, and the foundational systems that let product, finance, and operations teams make decisions. And they've become one of the most enthusiastic Claude Code adopters at Anthropic — discovering that AI-assisted workflows have fundamentally changed what's possible for a team of their size.
Here's a look at exactly how they use it.
Kubernetes Debugging With Screenshots
One of the most striking examples from this team involves infrastructure debugging — a domain where problems are often cryptic, error messages are dense, and resolution typically requires pulling in specialists who aren't always available.
When the team encountered pod IP address exhaustion in their Kubernetes cluster, the traditional path would have been: parse logs, search documentation, open a ticket with the networking team, wait. Instead, they fed screenshots of their dashboards directly into Claude Code and asked it to diagnose the issue.
Claude identified the root cause and returned the exact commands needed to resolve the problem — no networking specialists required, no waiting, no ticket queue. What would have been a multi-hour or multi-day incident was resolved in a single session.
This pattern — feeding visual context (screenshots, diagrams, dashboard captures) alongside log data — has become a repeatable workflow for the team. Claude's ability to reason across both visual and textual information makes it particularly effective for infrastructure problems where the relevant signal is often spread across different output formats.
Plain Text Workflows for the Finance Team
Perhaps the most democratizing use case from Data Infrastructure is what they've built for finance colleagues who have no coding experience.
The workflow is simple: finance team members write plain text files describing what they need. Something like: "Query the revenue dashboard for Q4 figures, pull the corresponding data from the transactions table, run a reconciliation against the ledger, and produce an Excel output with the variance highlighted."
That plain text file gets loaded into Claude Code. Claude reads the intent, figures out the implementation, and executes the entire workflow autonomously — querying dashboards, running SQL, formatting output.
The finance team doesn't write code. They don't configure pipelines. They describe what they want in the language they already use every day, and Claude handles translation and execution. This has removed a significant dependency on the data engineering team for routine reporting tasks and given finance colleagues genuine self-service capability.
Codebase Navigation for New Hires
Onboarding onto a large, complex data codebase has historically been one of the most friction-heavy parts of joining a data team. New data scientists often spend their first weeks just trying to understand which files matter, how pipelines connect, what the naming conventions mean, and who owns what.
The Data Infrastructure team has changed this by leaning on Claude Code as a guided codebase navigator. New hires use Claude Code to read CLAUDE.md files (which the team has invested in writing well), identify relevant files for their first tasks, and get explanations of data pipeline dependencies in plain language.
Instead of spending time figuring out the lay of the land, new team members can ask direct questions: "Which files handle the transformation for the revenue attribution pipeline?" or "What does this function do and what downstream jobs depend on it?" Claude answers immediately, with context pulled from the actual codebase.
This has effectively replaced the need for a traditional data catalog for many navigation tasks, and it has meaningfully accelerated how quickly new analysts can make productive contributions.
End-of-Session Documentation Updates
Documentation is one of those tasks that everyone knows is important and almost no one does consistently — because it competes directly with the work that feels more urgent. The Data Infrastructure team has found an elegant solution: make documentation an automatic byproduct of working sessions.
At the end of a Claude Code session, team members ask Claude to summarize what was accomplished and suggest improvements to the relevant CLAUDE.md files. Claude reviews the work done during the session, identifies gaps or outdated information in existing documentation, and drafts updates.
This creates a continuous improvement loop. The better the documentation gets, the better Claude Code performs in future sessions — because Claude uses those files to understand the team's context, tools, and conventions. It's a compounding feedback cycle that has made the team's codebase progressively easier to work with over time.
Parallel Task Management Across Repos
For engineers managing multiple projects simultaneously — which describes most senior data engineers — context switching is an enormous hidden cost. Getting back up to speed after leaving a task for hours or days can eat 30 minutes before you're actually productive again.
The Data Infrastructure team has addressed this by opening multiple Claude Code instances across different repositories, each maintaining its own full context. When a team member switches back to a project they haven't touched in two days, Claude Code still has the full context of what was happening: which files were being modified, what the goal was, what had been tried.
This is particularly valuable for longer-running investigations — like debugging a data quality issue that requires checking multiple upstream sources — where maintaining context across multiple sessions would otherwise require extensive note-taking.
Team Impact
- Resolved Kubernetes infrastructure issues without requiring networking specialists, turning multi-day incidents into single-session resolutions
- Finance team members with zero coding experience now execute complex data workflows independently
- New data analyst onboarding accelerated significantly — Claude Code replaces traditional data catalogs for codebase navigation
- The team can now monitor 200 dashboards for anomalies — a task that was previously impossible to do manually at that scale
- End-of-session documentation updates have created a compounding improvement loop for team knowledge
Top Tips From the Data Infrastructure Team
Write detailed CLAUDE.md files. This is the single highest-leverage investment you can make. The better you document your workflows, tools, naming conventions, and expectations in these files, the better Claude Code performs across every task. Treat them like onboarding docs for a very capable new team member.
Use MCP servers instead of BigQuery CLI for sensitive data. When working with sensitive datasets, MCP servers give you better security control, logging, and privacy protection than direct CLI access. The setup takes time upfront but pays dividends in auditability.
Share team usage sessions regularly. Hold periodic sessions where team members demo their Claude Code workflows to each other. The best workflows tend to spread organically once people see them in action — but only if there's a venue for that sharing to happen.