Claude Code and Claude Tag: team workflows explained

Thariq Shihipar, Sid Bidasaria, and Robert Boyce describe how their work with Claude Code has shifted toward giving it goals rather than supervising each tool call. Much of their discussion centers on Claude Tag, an AI agent in Slack that can use team conversations for product context. One participant says 70 to 80% of his work happens there, with the terminal interface or desktop app used for closer refinement.

The team explains how its coding assistant harness changes as models improve. They recall adding to-do lists around Sonnet 3.5 to help it finish longer tasks, then removing support they felt newer models no longer needed. Interactive HTML artifacts with diagrams and mockups also appear as an alternative way to gather user input.

For ongoing work, the speakers trace a move from laptop sessions to remote development boxes and hosted containers. They note that containers need access to the development environment, which takes setup. Routines can then revisit feedback and attempt fixes beyond a single session.

Code review supplies another example: parallel agents search for bugs, then adversarial reviews filter the findings for human attention. The speakers emphasize tests, screenshots, and feedback from usage data or Slack as checks on generated work.