Anthropic Developer Shares Fable 5 Prompting Tips to Uncover User Blind Spots

Anthropic developer Thariq Shihipar argues that with the advent of Claude’s latest model, Fable 5, the bottleneck in AI-assisted coding has shifted. The challenge is no longer primarily with the model’s capabilities but with the user’s own “blind spots”—specifically, their “unknown unknowns.” Shihipar suggests that Fable 5 is advanced enough that the quality of its output is largely determined by a user’s ability to clarify their own unarticulated knowledge gaps. This insight is particularly relevant for Indian developers and startups leveraging AI, as identifying these blind spots can significantly improve efficiency and project outcomes.

Understanding “Unknown Unknowns”

Shihipar categorises knowledge into “Known Knowns” (what’s in the prompt), “Known Unknowns” (questions one is aware of but hasn’t resolved), and “Unknown Knowns” (obvious knowledge rarely written down but recognised when seen). The most critical category, “Unknown Unknowns,” refers to aspects entirely unconsidered by the user. These can emerge deep within implementation or indicate that a problem requires a fundamentally different approach. According to Shihipar, effective agentic coders anticipate these unknowns, even if they have relatively few.

The Risk of Imbalance in Prompting

Too much specificity in prompting risks Fable 5 rigidly adhering to instructions, potentially missing more optimal solutions. Conversely, excessive vagueness can lead to decisions based on industry defaults that don’t align with the specific task at hand. Shihipar emphasises that failing to account for unknowns can lead to suboptimal results in both scenarios. However, Claude can assist in discovering these unknowns by rapidly searching codebases and the internet, leveraging its extensive knowledge base. The key lies in providing Claude with context about the user’s current thinking and experience level.

Techniques for Pre-Implementation Blind Spot Discovery

Shihipar outlines several techniques for the phase preceding actual code implementation. One such method is a “blindspot pass,” where Claude is explicitly asked to identify the user’s unknown unknowns. This is especially useful when working in unfamiliar parts of a codebase. For instance, a prompt might be: “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 and help me prompt you better.”

For areas rich in “unknown knowns,” such as visual design, Shihipar recommends brainstorming and prototyping. Instead of immediately implementing, he has Claude generate diverse design directions as HTML artifacts, allowing for iterative feedback. He often initiates coding sessions with an exploration or brainstorming phase to consciously define the project scope. Other techniques include structured interviews, where Claude systematically queries the user about ambiguities, prioritising questions that could impact architectural decisions. References, especially source code, are also crucial, as models like Claude Design can interpret underlying code rather than just visual screenshots.

Key facts:

Feature	Description
Model	Anthropic’s Claude Fable 5
Core Idea	User blind spots (“unknown unknowns”) are the new bottleneck in AI coding.
Techniques	Blindspot passes, structured interviews, prototyping, context-rich prompting.
Developer	Thariq Shihipar, Anthropic

Post-Implementation and Learning

Before actual work commences, Shihipar tasks Claude with creating an implementation plan that prioritises mutable components like data models and user-facing elements, leaving mechanical refactoring for later. During implementation, unknowns can still arise. He has Claude maintain a temporary “implementation-notes.md” file to track decisions. When unexpected edge cases appear, Claude is instructed to choose the conservative option, log the deviation, and continue working.

After implementation, Shihipar suggests two techniques: “pitches and explainers” (summary documents for stakeholders) and “quizzes.” In the quiz method, Claude generates an HTML report detailing changes with context, followed by a quiz. Shihipar only merges changes after successfully passing this quiz. He demonstrated these techniques during the creation of the Fable launch video, a new domain for him, by having Claude teach him about color grading to address his knowledge gaps. The more powerful AI models become, the more critical it is to invest time in defining unknowns and creating flexible implementation plans.

Source: The Decoder, https://the-decoder.com/anthropic-developer-shares-prompting-tips-for-fable-5-that-focus-on-finding-your-own-blind-spots-first/