Anthropic Guide Details Six-Step Process for AI-Driven Code Modernization

Anthropic's "Notes from the Field" series details a six-step framework for enterprises using Claude Code to modernize legacy systems. The guide argues that while agents accelerate code generation, the primary bottleneck shifts to mobilizing the organization around approvals and review. The process involves defining modernization targets (Uplift, Transform, or Reimagine), creating a "certificate" of checkable conditions for correctness, establishing a promotion policy for human review, setting prerequisites, building an agentic workflow, and running a pilot before scaling.

Defining Targets and Certificates

The guide recommends categorizing modernization into three types: Uplift (version updates), Transform (stack swap with fixed behavior), and Reimagine (new stack with new behavior). A key component is the certificate, a set of verifiable conditions such as passing test suites, performance bounds, and adversarial reviews by Claude, which serve as the basis for accepting changes. This ensures that human review can be scaled by relying on objective evidence rather than diff-by-diff inspection.

Promotion Policy and Prerequisites

The third step involves defining a promotion policy that dictates the depth of human review based on the change's blast radius and risk. The guide emphasizes involving business users and developers early to build consensus and ensure the certificate reflects their standards. Prerequisites include setting up CI/CD compatibility, test capacity, and a code-freeze policy if necessary. The guide notes that for complex "Reimagine" projects, the lack of an existing system to diff against makes the certificate more subjective and prone to variable outcomes.

Workflow and Cost Management

The final steps involve building a dynamic agentic workflow using the code modernization plugin and running a small-scale pilot to estimate token usage and costs. The guide suggests optimizing costs by using cheaper models for mechanical tasks and reserving advanced reasoning for complex logic. It also highlights the importance of using Claude to map dependencies and document workflows, which helps in understanding the legacy codebase before transformation.