Anthropic's Claude Code is helping software teams manage sprawling legacy codebases by treating the AI like a new developer who needs proper context and training. At the University of Washington's MacCoss Lab, researchers discovered that the same methodology used to onboard human developers to a 700,000-line codebase works remarkably well for AI, enabling teams to complete projects that had been abandoned for years.

What's the Real Problem With AI Coding Tools on Large Projects?

When Brendan MacLean, a Claude Developer Ambassador at MacCoss Lab, first tried using Claude through the browser interface, he ran into a familiar frustration. Each conversation felt like starting from scratch. Claude had no understanding of Skyline, the lab's open-source protein analysis software, or how its components related to each other. The AI couldn't grasp 17 years of accumulated development decisions and architectural patterns.

This wasn't a limitation of Claude itself. It was a context problem. MacLean realized he was facing the same challenge he'd tackled dozens of times before: how to bring new developers up to speed on a massive, complex codebase. The difference was that this time, the "new developer" was an AI tool.

"I could introduce Claude through Claude Code to my large project as I would a trainee developer: by explaining enough to achieve a successful limited project and produce improved context for the next iteration," explained Brendan MacLean, Claude Developer Ambassador at the University of Washington's MacCoss Lab.

Brendan MacLean, Claude Developer Ambassador at the University of Washington's MacCoss Lab

MacLean's insight transformed how his lab uses Claude Code. Instead of treating the AI as a magic tool that should understand everything instantly, he built a deliberate context layer that persists across sessions and grows over time.

How to Build a Sustainable AI Context Layer for Your Codebase

• Separate Context Repository: MacLean moved all AI context into its own repository called pwiz-ai, kept separate from the main codebase so it applies across all branches and development timelines. A CLAUDE.md file at the root handles environment setup and points Claude to relevant documentation, functioning as an orientation guide rather than a complete knowledge base.
• Encode Domain Knowledge in Skills: Skills are an open format for giving Claude capabilities and expertise. MacLean's debugging skill, for example, pulls Claude out of "guess and test" mode and pushes it toward root cause analysis before attempting any fix. Skills can be triggered manually or automatically, with critical ones tuned with explicit conditions.
• Use MCP Integrations for Real Data Access: Model Context Protocol (MCP) integrations allow Claude to access actual test results, exception reports, and support threads. MacLean built MCP servers in both C# and Python so Claude could access real data from LabKey Server, team email, and GitHub, enabling automated daily summaries of test failures and open support issues.

The results speak for themselves. A year-long project to build a Files View panel in Skyline, abandoned when the original developer left, was completed in two weeks with Claude Code. Three years ago, MacLean had stopped adding features to Skyline's nightly test management module after losing the developer who maintained it. Recently, with proper Claude Code setup, he spent less than a day adding features he'd wanted for years and updating page layouts with CSS.

How Is This Changing How Labs Actually Work?

The impact extends beyond individual projects. Screenshot reproduction for Skyline's 2,000-plus tutorial images is now fully automated and nearly 100 percent reproducible. Claude Code extended this capability with diff-only views and pixel change amplification, and even wrote the MCP server in C# to make it possible.

Perhaps most tellingly, developers in the lab are now spending less time writing code themselves and more time instructing Claude Code. One skeptical developer who had doubted agentic coding tools built and shipped a new plotting extension for visualizing ion mobility data, crediting Claude Code for the work. MacLean noted that he is seeing almost everyone taking on fun new features that they might have felt too buried in other work to attempt.

"Understand that Claude can't learn without you recording 'context.' Don't expect magic. Invest in building and maintaining your context layer. And treat it like any other project artifact: version it, grow it, maintain it," stated MacLean.

Brendan MacLean, Claude Developer Ambassador at the University of Washington's MacCoss Lab

The key insight MacLean emphasizes is that context is what persists across sessions, and it must be maintained deliberately. This is the part most developers skip, and it's why most developer success plateaus. The to-do lists and plans Claude generates don't carry over between conversations, but a well-maintained context layer does.

For teams managing legacy codebases, the implications are significant. Institutional knowledge that typically walks out the door when developers leave can now be encoded and maintained. Projects that would have been abandoned due to developer turnover can be revived. And the onboarding process for both human and AI collaborators becomes a deliberate, versioned artifact rather than an implicit understanding held by a few key people.

This approach represents a shift in how teams think about AI coding tools. Rather than expecting them to magically understand complex systems, teams are learning to treat context management as a core engineering practice, much like version control or testing. For Anthropic's Claude Code, this means the tool's real power emerges not from raw capability, but from how thoughtfully teams integrate it into their existing development workflows.