Summary
At the AI Engineer conference, Simon Willison sits down with Cat Wu and Thariq Shihipar — engineers at Anthropic — for an in-depth fireside chat on Claude Code and the newly launched Fable model. The conversation traces Claude Code’s arc from its February 2025 debut as a footnote on the Claude Sonnet 3.7 release to a full-blown agentic coding environment where engineers can “one-shot” entire features. The guests describe the day Fable went live — literally minutes before the session began — and the step-change it represents in autonomous task completion.
A centerpiece of the discussion is the revelation that the Claude Code system prompt has been cut by roughly 80%, driven by improvements in Opus 4.8 and Fable. The Anthropic team explains why: newer models are more creative and capable than the examples used to constrain them, and hard “do not do this” instructions increasingly conflict with user intent at runtime. Their updated philosophy favors sparse context over rigid rules, with evals proving that fewer instructions often produce better outcomes.
The guests also reflect on how agentic coding is reshaping the engineering profession — compressing six-to-twelve-month build cycles down to days and raising the premium on product taste and business judgment over pure execution skill. The conversation offers rare insider perspective on how Anthropic trains Claude to handle ambiguous intent, the pitfalls of over-specification in system prompts, and what “higher-quality work” means in a world where implementation is largely delegated to the model.
📺 Source: AI Engineer · Published July 15, 2026
🏷️ Format: Interview
