A Fireside Chat with Cat and Thariq from the Claude Code team

Main idea

Anthropic internally uses Claude Tag (a Slack integration of Claude Code) to generate 65% of all product engineering pull requests. The Claude Code system prompt for frontier models (Fable 5, Opus 4.8) shrank by 80% — adding examples to system prompts is counterproductive for frontier models, which prefer contextual guidance.

Context

The fireside chat took place at the AI Engineer World's Fair in July 2026. Cat Wu and Thariq Shihipar from the Claude Code team describe the evolution from January 2026 (internal auto mode deploy) to March (public release). Their product discipline: a feature ships externally only when employee retention demonstrates genuine sustained adoption.

Why it matters

65% PRs generated by AI is one of the first concrete numbers from real production AI deployment at a top AI firm — not marketing. For developers: shorter, more contextual system prompts work better for frontier models — a directly actionable insight for anyone designing AI agents.

Details / arguments

• "Adding examples to a system prompt is no longer best practice" for frontier models
• Prioritization: internal employee retention metrics decide what gets shipped externally
• Code review: critical changes still manual; peripheral layers increasingly rely on automated eval systems
• Auto mode security: Sonnet classifier evaluates permissions dynamically; red-teaming ran January–March before public release
• Tool cardinality: team keeps it "pretty low" so Claude can unambiguously distinguish when to call each tool