MacCoss Lab Applies 17 Years of Code Onboarding to Claude Code

TL;DR

• The MacCoss Lab at the University of Washington has successfully integrated Claude Code into their 17-year-old, 700,000+ line C# codebase for the Skyline protein analysis software.
• Leveraging established developer onboarding methodologies, they treated Claude Code like a new trainee developer, providing context and specialized skills.
• This approach helped overcome challenges with long-lived projects, such as reducing technical debt and accelerating development of complex features.
• The integration significantly sped up the development of a previously stalled Files View panel, completing it in two weeks with Claude Code as a co-author.

Brendan MacLean, a Claude Developer Ambassador and principal developer at the University of Washington's MacCoss Lab, has shared a unique approach to integrating AI coding tools into a mature software project. Their project, Skyline, an open-source protein analysis software, has been under active development since 2008. With a codebase exceeding 700,000 lines of C# and a history of contributions from numerous developers over 17 years, the lab faced challenges common to long-lived projects, including areas becoming difficult to maintain due to developer turnover.

Initially, MacLean was skeptical that modern AI coding tools could grasp the complexities of their extensive C# codebase. Early experiments with Claude.ai in the browser were limited, requiring him to describe problems without reference to the project's existing code. This made iterative development laborious, as each session felt like starting from scratch, with no inherent understanding of Skyline's architecture or its 17-year developmental history. This experience mirrored the challenges of onboarding new human developers to the project.

This parallel led MacLean to an innovative solution: applying the lab's proven developer onboarding methodology to Claude Code. He began by introducing Claude Code to the project as he would a trainee developer. This involved creating a dedicated repository, pwiz-ai, separate from the main Skyline codebase. This repository serves to establish context and provide Claude with the necessary information to understand the project's environment and documentation, acting as a foundational "lay of the land."

Within this pwiz-ai repository, MacLean implemented "skills," an open format for granting AI agents specific capabilities and expertise. For instance, a specialized debugging skill was developed to move Claude beyond a "guess and test" approach, encouraging root cause analysis before suggesting fixes. These skills can be triggered manually or automatically, with critical ones, like the debugging skill, being configured with explicit conditions to ensure they are loaded whenever specific issues arise, such as investigating bugs or unexpected behavior. This structured approach to context and skill development significantly reduced the overhead of teaching Claude the intricacies of debugging the codebase.

The impact of this methodology on development velocity and technical debt reduction has been significant. A notable example is the Files View panel, a new interface designed for Skyline to display all document-related files with file system monitoring and drag-and-drop organization. This project had been stalled for a year after the original developer left. MacLean picked it up using Claude Code, and within two weeks, the feature was completed, with all final commits co-authored by Claude. This success stands in contrast to past experiences where unfinished features were often abandoned due to the challenges of bringing new developers up to speed on complex, long-term projects. For developers interested in exploring Claude's capabilities for coding, resources are available on claude.com/blog/onboarding-claude-code-like-a-new-developer-lessons-from-17-years-of-development, and for those looking to integrate AI into their development workflows, the Claude Code plugin for Visual Studio can be found on the Visual Studio Marketplace. Developers can also explore the broader applications of Claude's AI agent capabilities at agentskills.io/home.

Summary

• The MacCoss Lab successfully applied its 17-year-old developer onboarding process to integrate Claude Code into a massive C# codebase, treating the AI as a new team member.
• By creating a separate context repository (pwiz-ai) and defining specialized "skills," they provided Claude Code with the necessary understanding to contribute effectively.
• This method significantly reduced technical debt and accelerated development, exemplified by the rapid completion of a previously stalled project feature in just two weeks.
• The approach highlights how established software development practices can be adapted to enhance the capabilities and integration of AI coding assistants in complex, long-lived projects.

Source: Onboarding Claude Code like a new developer: Lessons from 17 years of development | Claude