Inside Anthropic’s Claude Fable 5: How to Uncover Your Unknowns for Better Agentic Coding

Map vs. Territory

Thariq likens the prompt, skills, and context given to Claude to a map, while the actual codebase, real‑world constraints, and the work to be done represent the territory. The gap between them consists of unknowns . When Claude encounters unknowns it can only guess what you intend, and the more unknowns there are, the higher the chance of deviating from your true intent.

Four Quadrants of Unknowns

Known Knowns : Information explicitly stated in the prompt.

Known Unknowns : Gaps you are aware of and can ask for help.

Unknown Knowns : Assumptions you consider obvious but Claude cannot see.

Unknown Unknowns : Blind spots you haven’t realized exist.

Top agentic coders (e.g., Boris and Jarred) have relatively few unknowns because they deeply understand both the codebase and the model’s behavior.

Three‑Stage Methodology

Pre‑Implementation: Five Tools to Clear Blind Spots

1. Blind Spot Pass – When entering an unfamiliar domain, prompt Claude with the terms “blindspot pass” and “unknown unknowns” and describe your background.

Example: “I know nothing about the auth module of this codebase. Can you do a blindspot pass and list my unknown unknowns so I can give better instructions?”

2. Brainstorms and Prototypes – For ambiguous judgments (e.g., visual design), ask Claude to generate multiple directions at once.

Example: “I want a data dashboard but have no visual taste. Give me an HTML page with four completely different designs so I can react.”

3. Interviews – After brainstorming, let Claude interview you, prioritizing questions whose answers would change the architecture.

Example: “Interview me question by question, focusing on any ambiguous points that could alter the architecture.”

4. References – When you cannot describe what you want, provide concrete references.

Example: “The Rust crate `vendor/rate‑limiter` implements the backoff logic I need. Read it and reproduce the same semantics in our TypeScript API client.”

5. Implementation Plans – Before coding, ask Claude for an implementation plan that emphasizes the parts you are most likely to adjust (data model, type interfaces, UX flow) and pushes mechanical refactoring to the end.

Example: “Write an HTML implementation plan, putting the decisions I’ll likely change—data model, new types, user‑visible features—first; put pure refactoring at the bottom.”

During Implementation: Record Deviations

Because plans rarely survive unchanged, Thariq has Claude maintain an implementation-notes.md file that logs why each forced deviation occurs.

Example: “If you encounter an edge case and must deviate, choose a conservative solution, record it under a ‘Deviations’ section, and continue.”

Post‑Implementation: Ensure Understanding and Get Buy‑In

After shipping, the biggest challenge is approval. Package the prototype, spec, and implementation notes into an “explain artifact” that reviewers can consume, starting with a demo GIF.

Example: “Bundle the prototype, spec, and notes into a document I can drop into Slack for buy‑in, with the demo GIF up front.”

Use quizzes to verify you truly understand the changes before merging.

Example: “Give me an HTML report with change context, intuitive explanation, actions taken, and a quiz at the bottom that I must pass before merging.”

Real‑World Case: Producing a Release Video with Claude Fable 5

Thariq, not a video‑editing expert, asked Claude about Whisper transcription accuracy and whether ffmpeg could cut filler words. He then had Claude prototype the UI‑voice sync with Remotion.

When the final video appeared gray, he realized the issue was color grading—a domain he didn’t understand. Instead of asking Claude to generate variants, he let Claude teach him about color grading, exposing his unknowns.

The entire release video was edited by Claude‑generated code.

Final Takeaways

Stronger models increase the importance of the right approach. When a long‑running task fails, the problem is usually not the model’s intelligence but that you spent too little time defining unknowns or didn’t give Claude a flexible implementation plan.

Each explain, brainstorm, interview, prototype, and reference step helps discover unknowns cheaply before costly fixes become necessary. Before starting the next project, ask Claude Fable 5 to help you find your unknowns.