Mastering Claude Fable 5: How to Prompt by Finding Your Blind Spots

As AI models evolve from simple chatbots into sophisticated agentic coders, the bottleneck of productivity is shifting from model capability to user clarity. Anthropic developer Thariq Shihipar suggests that with the arrival of Claude's Fable 5, the limiting factor is no longer the AI's intelligence, but the user's ability to identify their own "unknowns."

The Four Quadrants of Knowledge in Prompting

To master Fable 5, Shihipar argues that developers must categorize their knowledge to avoid the pitfalls of poor prompting. He breaks down cognitive gaps into four distinct categories:

  • Known Knowns: The explicit instructions and data already present in your prompt.
  • Known Unknowns: Specific questions or gaps you are consciously aware you haven't solved yet.
  • Unknown Knowns: Intuitive knowledge so obvious you wouldn't think to write it down, but would recognize instantly upon seeing it.
  • Unknown Unknowns: The most dangerous category—critical elements, edge cases, or architectural requirements you haven't even considered.

When users fail to account for these, they fall into a trap: being too specific causes Fable 5 to follow instructions rigidly even when a better path exists, while being too vague leads the model to rely on industry defaults that may not fit the specific task.

Strategic Techniques for Uncovering Unknowns

Rather than jumping straight into implementation, Shihipar recommends a "blindspot pass." This involves using Claude to audit your mental model before a single line of code is written. For instance, if working in an unfamiliar codebase, you can prompt Claude to: "I'm working on adding a new auth provider but I know nothing about the auth modules in this codebase. Can you do a blindspot pass to help me figure out my relevant unknown unknowns?"

Other high-leverage techniques include:

  • Structured Interviews: Asking Claude to interview you question-by-question, prioritizing ambiguities that would fundamentally alter the system architecture.
  • Rapid Prototyping: Using HTML artifacts to have Claude generate radically different design directions, allowing you to react to visual "unknown knowns."
  • Implementation Planning: Directing Claude to create a plan that prioritizes volatile components like data models and type interfaces before moving to mechanical refactoring.

Closing the Loop: Documentation and Verification

The prompting process doesn't end once the code is generated. To manage unknowns that surface during execution, Shihipar suggests using Claude Code to maintain an implementation-notes.md file. This tracks every decision made, ensuring that if a task goes sideways, the reasoning is preserved.

To ensure complete comprehension before merging code, Shihipar employs a "quiz" method. Claude generates an HTML report detailing all changes and insights, followed by a quiz for the developer. He maintains a strict rule: do not merge the code until the quiz is passed without errors. This ensures that the developer has moved the project's "unknowns" into the realm of "knowns."

Key Takeaways

  • Focus on "Unknown Unknowns": The goal of advanced prompting with Fable 5 is to use the AI to reveal what you haven't even thought to ask.
  • Avoid the Specificity Trap: Balance your prompts to avoid being so rigid that the AI cannot improvise, or so vague that it defaults to generic solutions.
  • Implement a Verification Workflow: Use implementation notes, structured interviews, and post-implementation quizzes to turn AI-generated code into mastered knowledge.