How Anthropic Performs Large-Scale Code Migrations with Claude Code

Code migration is a project that ports a production codebase to a new language. It used to be a large-scale undertaking that took years, but now it's possible with AI.

Bun's Zig→Rust Migration: Bun co-founder Jarred Sumner used Claude Code to port one million lines of code to Rust in under 2 weeks. 100% of existing tests passed before the CI merge, and after the merge only 19 regressions occurred, all of which were fixed.

Python→TypeScript Migration of an Internal Tool: Anthropic Labs' Mike Krieger migrated a 165,000-line Python codebase to TypeScript over a single weekend. Compile speed dropped from 8 minutes to 2 seconds, and binary startup speed became 6x faster.

Core Insight: Rather than fixing the code, you need to fix the process (loop) that generates the code.

Why AI Makes Migration Possible:

• Work is parallelizable (can be handled independently by file or crate)
• Context is clear (existing code is the best spec)
• Objective means of verification exist (test suite)
• Compile errors/test failures automatically generate the next task
• Rule violations are detected immediately, so deviations don't accumulate

6-Step Migration Process: Step 1: Write a rulebook, dependency map, and gap list → Step 2: Stress-test the rules → Step 3: Translate everything → Steps 4·5·6: Repeatedly run the implement-review-fix loop

A prerequisite is a strong "judge." It must be a test suite capable of evaluating the original code and the ported code equally. If needed, rewrite tests so they're portable, and validate with an adversarial reviewer.

Cost: The Bun migration consumed roughly $165,000 based on API pricing (5.9 billion input tokens, 690 million output tokens). A project that used to take years is now possible in a matter of weeks, at a cost in the hundreds of thousands to millions of dollars.