Anthropic: With Fable 5, the Bottleneck Is the User, Not the Model

Anthropic developer Tariq Shihipar published a breakdown of a new approach to working with Fable 5 — the latest Claude model. His central argument: output quality is no longer determined by the model's raw power but by the user's ability to identify their own blind spots before work begins. "Instructions that are too detailed force the model down the wrong path; too vague, and it produces generic solutions," Shihipar explains.

1Three Key Techniques

The first is the "blindspot pass": before defining the task, ask Claude to surface unknown unknowns in your understanding of the problem. The second is structured interviews, where the model formulates its own questions and ranks them by their impact on the solution's architecture. The third is post-implementation self-check quizzes: Claude writes a test to confirm that the developer actually understood what was built.

As an example, Shihipar describes building a video editor: before writing any code, he worked through Whisper, ffmpeg, and the basics of color grading — only then did he hand off the implementation to Claude. This is a fundamentally different workflow from how most people use AI assistants today: understand it yourself first, then delegate.