Anthropic dev: Claude Fable 5 quality now limited by user's blind spots, not model

Key takeaway

• Anthropic developer Thariq Shihipar argues that Claude Fable 5's quality is now limited by users' blind spots rather than the model itself.

• He describes a systematic approach—including brainstorming, prototyping, and structured interviews before implementation—to help developers uncover what they don't know they don't know.

• This reflects a broader shift in AI-assisted development: as models become more powerful, success depends less on the tool and more on the user's ability to clarify ambiguities and anticipate edge cases.

3 Key Points

FAQ

What are the four categories of knowledge Shihipar describes?

Known Knowns (what is stated in the prompt), Known Unknowns (questions you know you haven't answered), Unknown Knowns (obvious knowledge you wouldn't write down but would recognize), and Unknown Unknowns (things you haven't considered at all). Unknown Unknowns are the critical category that limits output quality.

What techniques does Shihipar recommend before starting implementation?

A blindspot pass (asking Claude to identify unknown unknowns), brainstorming and prototyping (especially for areas like visual design), structured interviews where Claude asks prioritized questions, using source code as references, and creating an implementation plan that focuses on parts most likely to change such as data models and type interfaces.

Why is being too specific in prompts a problem?

Too much specificity risks Fable 5 rigidly following instructions even when a change of course would make more sense, whereas too much vagueness results in decisions based on industry defaults that don't fit the specific task.