Tracking Its Unknowns: An Anthropic Engineer's Method for Claude Fable

Anthropic engineer Thariq Shihipar outlines a method for Claude Fable in Claude Code to help users map blind spots and pilot advanced AI coding agents.

How do you pilot a code agent when the model becomes more capable than the instructions you give it? Thariq Shihipar, an engineer at Anthropic, published a methodological guide on the company's blog dedicated to working with Claude Fable in Claude Code. His thesis: the quality of the result is no longer limited by the model but by the user's ability to clarify their own blind spots. He classifies these into four categories, from what is explicitly known to what is unknown unknowns, including what is known but never formulated.

The text details techniques applicable before, during, and after implementation: asking Claude for a "blind spot pass" to map the blind spots of an unfamiliar domain, prototyping several directions in HTML before writing production code, being interviewed question by question about project ambiguities, pointing to existing source code as a reference rather than a screenshot, having the agent keep a log of decisions made along the way, then being quizzed on changes before validating anything.

The author illustrates the approach with the launch video for Fable, which he claims was edited end-to-end with Claude Code, including transcription and color grading—a domain he admits he discovered by being trained by the model itself.