Software engineers are using large language models to manage technical debt in legacy codebases. Technical debt accumulates when code is modified quickly to meet business needs, and cleanup budgets are often limited. Traditional cleanups require significant time spent on decisions and implementation. Large language models struggle with brownfield projects because they lack a consistent understanding of terminology and context. The models guess incorrectly when the codebase contains multiple definitions of the same concept. The work is divided into strategic planning, which defines the needed changes, and tactical execution, which implements those changes. Strategic work requires analysis of the system and alignment with business goals. Agents are coordinated through a skill system that loads predefined markdown instructions and runs sub‑agents with focused contexts. Each sub‑agent reports results, enabling incremental reviews and pull requests. Each repository contains a .workflow.json file that declares the domain, bounded contexts, and relationships between them. This manifest is generated from domain blocks and a living glossary per context. The model uses the map to determine which context it is operating in and to resolve terminology disputes. Discrepancies are reported as issues that are assigned to the owning repository, then addressed by an agent and reviewed before merging. Migrating a context to a true domain model clarifies ownership and eliminates guesswork. The system is ready for broader adoption and will be updated with additional skills. Readers are invited to subscribe for further details.