Claude Code Team Dialogue: Auto Mode as Safety Foundation, System Prompts Cut by 80%

In a recent AI Engineer World's Fair fireside chat, Anthropic's Claude Code team members Cat Wu and Thariq Shihipar shared key practices. Claude Tag (Slack collaboration integration) now submits 65% of PRs for the product engineering team, reflecting a shift in collaboration. System prompts have been reduced by 80% thanks to model improvements (e.g., Fable 5 and Opus 4.8); the team found that reducing examples and "don't" instructions actually improves performance. Auto mode has been used internally for months, extensively red-teamed and eval-validated, and is seen as crucial for safety—especially for tools like Claude Tag that operate in public channels. Thariq noted that models excel at rewriting code, and the team has rewritten Bun in Rust for internal use. Cat emphasized that human review remains for core changes, but peripheral code has been gradually automated, with trust built through continuous iteration of review tools.