A 700,000-line C# codebase. The core developer graduated and left. Feature modules sat frozen for three years, too risky for anyone to touch. Brendan MacLean, lead developer at the MacCoss Lab at the University of Washington, used Claude Code to finish this abandoned build in two weeks. His method wasn't making the AI smarter — it was giving the AI a complete context system, the way you'd bring up an intern. This article breaks down exactly how he did it, alongside a comparison with Symphony, the parallel orchestration scheme OpenAI shipped the same day.

The Core Idea: Bring Up the AI Like an Intern

Brendan's insight is plain: every time a newcomer joins the lab, they don't know the project's full picture, how the code fits together, or the unwritten rules. The difference is that a newcomer learns over time, while Claude forgets everything the moment a conversation ends.

So his core strategy: turn tacit knowledge into explicit, AI-readable assets.

The Three-Layer Context System

Layer 1: CLAUDE.md — the Project's Terrain Map

Brendan set up a separate repository, pwiz-ai, dedicated to context written for the AI to read. The CLAUDE.md at the root is the 'terrain map,' telling Claude:

• What the project is and what it's for
• The codebase structure
• How to build it
• The testing workflow

Project address: github.com/ProteoWizard/pwiz-ai

The key principle: CLAUDE.md solves 'knowing where things are.' It does not solve 'knowing how to do the work.'

Layer 2: Skills — the Domain Knowledge Base

The real domain expertise lives in skills. Brendan wrote a debugging skill, for example, described as 'always load when investigating bugs, failures, or unexpected behavior' — it forces Claude to do root-cause analysis before acting, instead of blind guessing and poking.

Layer 3: MCP — Let the AI See Real Data

Through MCP (Model Context Protocol) integration, Claude can read real test data, exception reports, and user tickets. With all three layers stacked, every interaction starts from 'already understands' instead of 'blank slate.'

Results in Practice

With the context in place, Brendan began clearing years of accumulated technical debt:

• A three-year abandoned module: a test-management module written in Java, on a stack completely different from the main C# codebase. Done with Claude Code in a day.
• 2,000+ tutorial screenshots: once maintained entirely by hand, now fully automated and nearly 100% reproducible.
• A daily automatic report: generated by Claude, summarizing overnight test failures, exceptions, and unresolved tickets.
• A brand-new feature dashboard: the developer in the lab least interested in AI coding tools built and shipped a brand-new data-visualization dashboard with Claude Code.

OpenAI's Answer: Symphony

The same week, OpenAI released the open-source project Symphony (over 18,000 stars on GitHub already).

Project address: github.com/openai/symphony

Symphony takes a completely different approach: it turns a Linear project board into the control center for AI coding. Every Issue in the Open state gets an Agent assigned automatically; the Agent keeps running in its own workspace and restarts automatically if it crashes.

In some teams' first three weeks on Symphony, the number of successfully merged PRs surged 500%.

Choosing Between the Two Routes

Shared finding: what used to travel by word of mouth and muscle memory now has to be written down in black and white before the AI can carry it. Whichever route you take, turning tacit knowledge into explicit assets is a step you cannot skip.

What You Can Do Now

If you're coding with AI assistance, three things worth doing today, whatever the tool:

1. Write a CLAUDE.md (or your equivalent context file): project structure, build commands, test workflow, red-line rules — spelled out line by line.
2. Turn common operations into Skills: debugging procedures, code-review checklists, deployment steps — so the AI never starts from zero.
3. Keep the docs current: code on iteration 7 or 8 with docs still at version 1.0 is far too common. Stale docs are more dangerous than no docs.