Anthropic outlines steps to prepare for AI-driven code modernization

The guide then describes how to build a "certificate" of evidence that each change must satisfy, involving developers, user groups, and business leads in its design. A tiered promotion policy is recommended to balance speed and review depth, with lighter human review for high‑risk, time‑critical updates. Claude Code and the code‑modernization plugin are suggested for automating discovery, rule extraction, and token‑usage measurement, while models like Sonnet can handle high‑volume mechanical checks and more capable models can address complex transformations. The article also advises measuring token costs on a pilot, extrapolating to estimate a cost floor, and optimizing the workflow to reduce expensive retries.