How We Ported 900,000 Lines of Code in 4 Days Using the Ralph Wiggum Loop

In four days, our Rōnin co-founder, Byron McClain, ported a 900,000-line video game from Windows to macOS, without writing a single line of code himself.

The game? Command & Conquer: Red Alert, the 1996 real-time strategy classic from Westwood Studios.

The experiment? Take an open-source game engine built exclusively for Windows, completely rebuild it to run natively on macOS, and do so entirely using Claude Code, an AI coding agent, paired with the Ralph Wiggum Loop methodology.

When we asked our co-founder, Byron McClain, why he chose to test AI coding agents on a 28-year-old video game, his answer was simple: “I’m not a game developer. I know Jack about how to develop a game.”

And that was exactly the point.

“I wanted to find the edges of Claude Code,” Byron explains. “I needed something gigantic and pre-existing that I could try this on, but it would need to have very clearly defined goals. I don’t know how to do any gaming, anything. So I thought, here’s a perfect test: take something I’ve never done before and see what Claude Code paired with the Ralph Wiggum Loop could actually handle.”

Turns out, it can handle quite a bit.

Using Claude: greenfield vs. brownfield

When testing new AI coding tools, most people start with greenfield projects, or building something new from scratch. It’s easier. There’s no existing code to untangle, no legacy dependencies, or any architectural decisions made 30 years ago that you have to work around.

At Rōnin, we have already tested Claude Code on greenfield work. It performs well. You can take a product requirements document, feed it into Claude, and watch Claude generate a functioning piece of software phase by phase.

But brownfield projects? That’s where things get interesting. With a preexisting codebase, you need some prior knowledge of what you’re looking at. A developer must understand context and navigate and update code they didn’t write. This is exactly the challenge Byron was looking for.

The Red Alert engine was the perfect stress test. At over 900,000 lines of code, it represented decades of Windows-specific development. Every system dependency would need to be identified, isolated, and replaced with macOS equivalents. The windowing system, graphics rendering, audio processing, file system access, network multiplayer protocols, all of it had to be rebuilt.

And Byron wanted to try it in Rust, a modern systems programming language, to prove Claude could handle not just porting but modernization.

Enter the Ralph Wiggum Loop

Claude Code has a feature called “plan mode” where it scans your entire codebase, asks clarifying questions, and generates a structured approach to whatever you’re trying to accomplish. “Having Claude in plan mode to plan out all the phases is extremely helpful when tackling something this massive,” Byron explains.

But plan mode alone wasn’t enough for a project this size, because even with these planned steps, someone still had to babysit Claude through execution. A human needed to answer questions, approve each step, and tell it when it was okay to move to the next phase. For a 900,000-line project? That would mean constant human intervention.

But Byron found another way, by means of the Ralph Wiggum Loop. Which is essentially “danger mode” for Claude Code and a way to run the AI agent autonomously without constant human approval.

Here’s how it Ralph Wiggum works: You break your project into phases, then break each phase into discrete tasks with clear acceptance criteria. Claude runs through each task, checking whether the criteria are met. If the criteria are met, it moves to the next task. If not, it keeps iterating until it gets it right or until it hits a predefined escape hatch and stops.

The human’s role is now to check in every 30 minutes or so, see where things stand, and kick off the next phase when one is complete.