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

Simon Willison · Simon Willison · 2026-07-21

Simon Willison publishes an annotated transcript of his fireside chat with Anthropic Claude Code team leads Cat Wu and Thariq Shihipar, revealing that Claude Tag lands 65% of Anthropic's product engineering PRs and that the Claude Code system prompt shrank by 80% for frontier models.

Open original ↗

Appears in

AI-Assisted Coding Culture: Landmark Rewrites, PR Description Backlash, and Prompting Debates

Extraction

Topics: coding-agentsprompt-engineeringagentic-aiai-safetyanthropic-internal-culture

Claims

Claude Tag, Anthropic's collaborative Slack-native coding agent, now lands 65% of product engineering pull requests for the Claude Code team internally.
The Claude Code system prompt was reduced by 80% for frontier models like Fable 5 and Opus 4.8 because providing examples and long prohibition lists now degrades rather than improves output quality.
Anthropic's auto mode uses a real-time Sonnet classifier that evaluates every tool call and its conversational context to enforce dynamic permissions and block prompt-injection attacks.
Critical core changes to Claude Code still require manual code-owner review, but outer-layer changes are increasingly reviewed only by an automated AI code-review GitHub bot.
Rewrites of entire codebases are now considered viable and even preferable given AI assistance, reversing the conventional software-engineering wisdom against rewrites.
Claude Tag's channel-specific team memory is implemented as a per-channel markdown file.

Key quotes

Claude Tag currently lands 65% of our product eng PRs right now. And this is a huge shift; this is more than 50% of our PRs.

Adding examples to a system prompt is no longer best practice for models like Fable 5 or even Opus 4.8. The Claude Code system prompt recently reduced in size by 80%.

The way you offset that is by being more ambitious.

Whenever you give a prompt to the model, you should think about the ways in which it could be misinterpreted by a well-intentioned human, in order to better understand how the model might interpret it.