Anthropic dev: Fable 5 prompts fail from your blind spots

Key Takeaways

• Fable 5's output quality is limited by the user's ability to identify their own blind spots, not model capability
• Being too specific locks the model into rigid instructions; too vague leads to generic defaults
• Systematic 'blindspot passes' and structured interviews with Claude can surface unknown unknowns before implementation

Anthropic developer Thariq Shihipar argues that with Fable 5, prompt failures stem less from model limitations and more from the user's own blind spots. In a detailed breakdown, he outlines techniques for uncovering what he calls 'unknown unknowns' before you write a single line of implementation code.

Why your unknowns matter more than your prompts

Shihipar borrows from the Rumsfeld matrix to categorize what users bring to a prompt. 'Known Knowns' are what you've explicitly stated. 'Known Unknowns' are questions you're aware you haven't answered. 'Unknown Knowns' are things so obvious you'd never write them down but would recognize immediately. The critical category is 'Unknown Unknowns,' the things you haven't considered at all.

His core claim: Fable 5 is the first model where output quality is limited by the user's ability to clarify their unknowns. The model itself is capable enough. The bottleneck is now human.

The specificity trap

Shihipar warns that over-specifying your prompt is just as dangerous as being vague. Too much detail and Fable 5 rigidly follows instructions even when a change of course would make more sense. Too little and you get decisions based on industry defaults that don't fit your task.

"When you don't account for your unknowns you fail both ways," he writes. The best agentic coders, he argues, have relatively few unknowns but still expect them.

Techniques for uncovering blind spots

Shihipar describes several pre-implementation techniques. The first is a 'blindspot pass,' where you explicitly ask Claude to identify your unknown unknowns. He offers this example prompt:

“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.”

— Thariq Shihipar, Anthropic

For visual design, where 'unknown knowns' dominate, he recommends brainstorming and prototyping. Instead of jumping to implementation, have Claude generate several radically different design directions as HTML artifacts. You react to them rather than describing them upfront.

Another technique: structured interviews. Claude asks the user question by question about ambiguities, prioritizing questions whose answers would change the architecture. Shihipar says he starts almost every coding session with an exploration or brainstorming phase to consciously define project scope.

References and implementation planning

Source code is the best reference, Shihipar says, even if it's in a different programming language. Claude Design reads a website's underlying code, not just the screenshot.

Before the actual work begins, he has Claude create an implementation plan that focuses on the parts most likely to change: data models, type interfaces, and everything on the user side. Mechanical refactoring comes last.

During and after implementation

Unknowns surface during implementation too. Shihipar asks Claude Code to keep a temporary 'implementation-notes.md' file where it tracks decisions so both parties can learn from the next attempt. When unexpected edge cases come up, Claude should pick the conservative option, log the deviation, and keep working.

After implementation, he recommends two techniques. First, 'pitches and explainers,' summary documents for stakeholders that bundle the prototype, specs, and implementation notes. Second, 'quizzes,' where Claude generates an HTML report detailing the changes made, with context and insights, followed by a quiz. Shihipar says he doesn't merge until he passes the quiz without errors.

A practical test: editing a launch video

Shihipar demonstrates these techniques with the Fable launch video, which he edited entirely with Claude Code. Video editing was new territory for him. By systematically uncovering his blind spots before starting, he was able to complete the project despite having no prior experience in the domain.

Frequently Asked Questions

What is a blindspot pass in AI prompting?

A blindspot pass is a technique where you explicitly ask the AI to identify your unknown unknowns before starting implementation. You describe your current understanding and ask the model to surface what you might be missing.

Why does being too specific hurt AI prompts?

Over-specification can lock the model into rigid instructions, causing it to follow your plan even when a different approach would work better. The model has no room to course-correct.

What are unknown unknowns in the context of AI coding?

Unknown unknowns are things you haven't considered at all. They're different from known unknowns, which are questions you know you haven't answered yet.

How do you document decisions during AI-assisted implementation?

Shihipar recommends having Claude Code maintain a temporary implementation-notes.md file where it tracks decisions and deviations. This creates a learning log for future attempts.

Source: The Decoder / Matthias Bastian