A Fireside Chat with Cat and Thariq from the Claude Code team
- source_url: https://simonwillison.net/2026/Jul/21/cat-and-thariq/
- source_type: article
- platform: blog
- author: Simon Willison; Cat Wu, Thariq Shihipar (Anthropic)
- original_date: 2026-07-21
- added_date: 2026-07-22
- category: coding
- tags: claude-code, claude-tag, system-prompt, auto-mode, anthropic, product
- quality_score: 5
- youtube: https://www.youtube.com/watch?v=uU5Gv2h8-9g
摘要（中文）
Simon Willison 与 Claude Code 团队 Cat Wu、Thariq 在 AI Engineer World's Fair 对谈纪要。硬信号：Claude Tag 已贡献团队约 65% 产品工程 PR；功能先对 Anthropic 员工开放，看 retention 再外放；关键改动仍人工审，外层更多自动评审；新模型上 system prompt 示例堆法与长 don't 清单有害，system prompt 近期约缩 80%；团队看好 auto mode 作为 Claude Tag 底座；内部 dogfood 称 ant fooding。对 Agent 产品 checklist 直接可用。
Summary (English)
Edited transcript of Simon Willison's fireside chat with Cat Wu and Thariq Shihipar from Anthropic's Claude Code team. Highlights: Claude Tag lands ~65% of the Claude Code team's product-engineering PRs; features ship to Anthropic employees first and only externalize what retains; critical changes still get human review while outer layers lean on automated review; for models like Fable 5 / Opus 4.8, stacking examples and long 'don't do X' lists can hurt—Claude Code's system prompt recently shrank ~80%; the team strongly believes in auto mode as enabling tech for Claude Tag. High-density product signals for coding-agent builders.
One-liner
Claude Code 团队：Tag 吃 65% 产品 PR，system prompt 砍约 80%，先看内部 retention。
Source body / metadata
Fetched via opencli web read / official docs during evening intake. Key claims are grounded in the source page and same-day Obsidian digest notes.
Edited transcript of Simon Willison's fireside chat with Cat Wu and Thariq Shihipar from Anthropic's Claude Code team. Highlights: Claude Tag lands ~65% of the Claude Code team's product-engineering PRs; features ship to Anthropic employees first and only externalize what retains; critical changes still get human review while outer layers lean on automated review; for models like Fable 5 / Opus 4.8, stacking examples and long 'don't do X' lists can hurt—Claude Code's system prompt recently shrank ~80%; the team strongly believes in auto mode as enabling tech for Claude Tag. High-density product signals for coding-agent builders.
Obsidian evidence
- local_note: OpenClaw定时任务/AK-RSS-Digest（89源精选）/2026-07-22-AK-RSS-Digest.md
- intake_run: daily-intake-evening 2026-07-22