Felix Pinkston
Apr 28, 2026 17:04
MacCoss Lab’s progressive use of Claude Code reworked its method to managing a 700,000-line legacy codebase, accelerating growth and decreasing tech debt.
MacCoss Lab, primarily based on the College of Washington, has spent 17 years sustaining Skyline, an open-source software program device used for protein evaluation. With over 700,000 traces of C# code and over 200,000 automated nightly assessments, the codebase is a behemoth that has challenged generations of builders. However Brendan MacLean, principal developer and Claude Developer Ambassador, discovered a novel strategy to handle this legacy: treating Claude Code, an AI-powered coding device, as he would a brand new developer.
Skyline’s longevity means its codebase carries a long time of collected technical debt. Builders rotating out and in usually left partially accomplished initiatives or untouched code areas. In line with MacLean, onboarding new builders was essential to maintain the challenge purposeful—and now that very same methodology is being utilized to AI instruments.
AI as a “Trainee Developer”
Initially skeptical about whether or not Claude Code might deal with the nuances of Skyline’s complicated codebase, MacLean examined it by isolating small issues. The outcomes had been underwhelming. Each interplay with Claude felt like ranging from scratch because of the lack of challenge context. However this sparked an thought: What if he onboarded Claude as if it had been a brand new developer?
To realize this, MacLean created a separate repository, pwiz-ai, to accommodate all AI-related context. A rigorously maintained CLAUDE.md file gives an outline of the challenge surroundings, whereas particular person “expertise”—task-specific capabilities—assist Claude deal with points systematically. For instance, a debugging ability prompts Claude to concentrate on root trigger evaluation as an alternative of trial-and-error fixes.
With this construction, Claude began contributing meaningfully. A protracted-abandoned challenge to create a Information View panel in Skyline was accomplished in simply two weeks, with remaining commits co-authored by Claude. MacLean famous related success in updating Skyline’s nightly take a look at administration module, which had sat untouched for 3 years after the unique developer left.
Remodeling Growth Workflows
Claude Code’s impression at MacCoss Lab goes past finishing unfinished options. The lab now makes use of the device to automate tedious duties like regenerating Skyline’s 2,000+ tutorial photographs and creating day by day error summaries. MacLean even credit Claude with writing an MCP (Message Management Protocol) server in Python to unify knowledge streams from numerous sources, enabling a centralized abstract of take a look at failures and help points every morning.
One of many lab’s builders, initially skeptical of AI instruments, efficiently constructed a mobilogram pane for visualizing ion mobility knowledge. MacLean says the device has allowed builders to tackle initiatives they beforehand prevented on account of time constraints or complexity.
Recommendation for Managing Legacy Codebases
MacLean’s expertise gives priceless classes for builders grappling with ageing codebases:
- Context is vital: Keep an in depth context repository, separate from the principle codebase if wanted, to make sure continuity throughout branches and developer turnover.
- Construct a ability library: Use AI expertise to encode area information and task-specific directions. Hold these light-weight and simple to take care of by linking to central documentation.
- Leverage MCP integrations: When AI instruments want real-time entry to knowledge, construct integrations to unify numerous knowledge streams. This method allowed MacLean’s lab to automate workflows and enhance developer effectivity.
A Mannequin for Open Supply Tasks
MacLean’s method has broader implications, particularly for open-source initiatives the place institutional reminiscence is scarce. By investing in a structured context layer, initiatives can guarantee continuity and scalability, whilst contributors come and go. The pwiz-ai repository itself is open supply, designed to profit the challenge and its contributors over the long run.
MacLean’s key takeaway? Treating AI as a trainee developer—with correct onboarding and context—can unlock its potential in ways in which go far past easy code technology. For groups managing sprawling legacy codebases, this system might be a game-changer.
Picture supply: Shutterstock
