Jeffrey Lee-Chan demonstrates a workflow in which short messages become coding tasks through a manager that retains goals, specifications and conversation history. Separate workers focus on repository implementation. Terminal sessions and parallel worktrees make delegated work visible, but the setup still requires human supervision and debugging.
Jeffrey Lee-Chan argues that the manager should assess evidence rather than adopt a worker's claim of success. He describes an independent skeptical reviewer, automated tests and browser checks. These are proposed controls for unreliable output, not proof that the demonstrated system always completes work correctly.
Jeffrey Lee-Chan uses an AI role-playing application and a multi-model answer comparison application as examples of harness design. The narrated demonstrations are not independent code or performance audits. He also distinguishes execution isolation from a separate staging environment, with integration testing needed before promotion; his staging setup remains unfinished.
Jeffrey Lee-Chan closes with operating limits: parallel work can consume more tokens, quotas can force model changes, and coordination does not eliminate costs or failure. His central architectural tradeoff is preserving intent while keeping implementation context focused. Model choices should fit the task and budget, not a universal ranking.
Watch on YouTube
