A field guide to Claude Fable 5: Finding your unknowns

URL: https://claude.com/blog/a-field-guide-to-claude-fable-finding-your-unknowns
Published: 2026-07-06
Captured: 2026-07-10 07:00 KST
Fetch method: ultimate-fetcher --json via Jina.

Summary

The article frames agentic coding as the practice of managing unknowns. A user arrives with known knowns, known unknowns, unknown knowns, and unknown unknowns. Strong agentic coders have fewer unknowns because they know the codebase, the model behavior, and their own quality criteria. But they also deliberately assume that unknowns exist and design the collaboration with Claude to expose them earlier.

The core recommendation is to help Claude help you. If a prompt is too specific, Claude may follow a bad path even when a pivot is needed. If it is too vague, Claude may fill gaps with generic best practices that do not match the task. The article’s patterns help users expose blind spots before implementation, capture new context during implementation, and test understanding after implementation.

Unknowns taxonomy

The article uses four categories:

• Known knowns: what is explicitly in the prompt.
• Known unknowns: what the user knows they have not figured out.
• Unknown knowns: criteria or taste the user recognizes when they see it but has not verbalized.
• Unknown unknowns: gaps the user has not considered at all.

Reducing and planning for unknowns is presented as a core skill of agentic coding.

Pre-implementation patterns

Blind spot pass

When entering an unfamiliar codebase, domain, or design area, ask Claude to find relevant unknown unknowns. The literal phrase “blind spot pass” is recommended. The prompt should include the user’s starting point and experience level so Claude can calibrate its explanation.

Example: ask Claude to inspect a codebase area and identify questions, assumptions, historical pitfalls, and better prompts before implementation starts.

Brainstorms and prototypes

Unknown knowns often appear in design and product work. The user may not be able to define the desired result in advance but can recognize it when they see it. In these cases, the article recommends asking Claude for several approaches or prototypes before committing to a full implementation. This makes implicit criteria visible while changes are still cheap.

Interviews

Claude can interview the user to elicit criteria and hidden constraints. This is useful when the user has tacit knowledge but has not turned it into a prompt. The interview transforms unknown knowns into known knowns.

References

References help Claude match the user’s mental model. Examples, screenshots, previous implementations, docs, and external resources reduce ambiguity and make taste or precedent more concrete.

Implementation plans

Before coding, Claude should produce a plan that surfaces assumptions, options, risks, and verification steps. The plan is not just a to-do list; it is a mechanism for exposing unknowns early.

During implementation

Implementation notes

The article recommends keeping implementation notes during the coding session. These notes preserve decisions, surprises, failures, and follow-up questions. They become a running context ledger that helps Claude and the user maintain shared state instead of relying only on chat history.

Post implementation

Pitches and explainers

After implementation, ask Claude to explain or pitch what was built. A good explainer reveals whether the work has a coherent rationale and whether the agent understood the intended value.

Quizzes

Ask Claude to quiz the user or itself about the implementation. Quizzes reveal whether important concepts and constraints were internalized. They also uncover gaps that might otherwise remain hidden until later maintenance.

Fable launch example

The article closes by showing how these patterns were used while launching Fable. The workflow moves from mapping unknowns, to exploring prototypes, to maintaining notes, to explaining and testing understanding. The point is to make the map and the territory converge: the prompt, plan, implementation, and user understanding should refer to the same reality.

Operational takeaway

Agentic coding quality depends less on a single perfect prompt and more on a process that continuously converts unknowns into explicit context. Claude is valuable not only as a code writer but as a collaborator that can search, prototype, interview, document, and test understanding.