Keeping AI-Assisted Projects Honest Across Multiple Sessions

The article presents a methodology for maintaining architectural integrity in complex AI-assisted software projects that span multiple development sessions.

Details

The author faced the challenge of maintaining architectural consistency in a complex AI-assisted project that spanned multiple development sessions. To address this, they introduced a methodology that splits the work across three roles: the Project Owner (the human), the Supervisor (a long-lived Claude chat session), and the Executor (Claude Code). The Supervisor holds the project history, makes architectural decisions, and translates them into precise prompts for the Executor. The Executor, in short-lived sessions, implements the plans without making high-level judgments. To compensate for the Executor's lack of long-term memory, the team maintains three persistent documents: ARCHITECTURE.md, DECISIONS.md, and CLAUDE.md, which act as the project's external memory. The process involves an initial setup phase to establish the foundation, followed by an iterative development loop where the Supervisor and Project Owner collaborate to refine the project.