How the Claude Code Team Uses AI to Build Claude Code

Summary

A video interview with the Claude Code team describes how its engineering practices have changed as model capabilities improve rapidly. The team says Claude Tag, an AI agent native to Slack, now handles 70% to 80% of its routine work, whereas engineers a year earlier reviewed individual tool calls, parameters, and reasoning decisions. Instead of supervising each step, they increasingly give the agent a goal and let it determine how to complete it, while Slack hides the model’s internal reasoning from users. The team also uses Claude Tag to develop itself: it identifies relevant stakeholders, creates prototypes and implementations, adds instrumentation, deploys tools internally, monitors usage, and alerts owners when feedback suggests changes. The engineers argue that trust can matter more than constant supervision in this workflow. Because model capabilities can undergo fundamental shifts roughly every two months, features built to compensate for earlier limitations may quickly become obsolete; the article cites the removal of a to-do-list function after improved multi-step memory and the shift from AskUserQuestion toward visual artifacts. The team therefore favors composable primitives such as permissions, visualization, verification, code review, and feedback over fixed solutions. Its work has also moved from inspecting tokens and tool calls to managing sessions, goals, and persistent systems running across conversations and cloud environments. Code review is increasingly organized through broad issue discovery, adversarial checks, cross-review, and multi-agent workflows, while humans focus on the problems that require attention.