@trq212 - Claude Fable 5: Finding

Open original

@trq212 on X A practitioner's guide to working with Claude Fable 5 by mapping what you know, don't know, and haven't considered before and during agentic coding work. The core idea is that output quality tracks how well you clarify the unknowns Claude runs into.

  • Fable is described as the first model where output quality is bottlenecked by the user's ability to clarify unknowns, not by the model itself.
  • The author sorts unknowns into four buckets: known knowns, known unknowns, unknown knowns, and unknown unknowns, and suggests prompting for each (e.g., a "blindspot pass" for unknown unknowns).
  • Techniques span the whole workflow: brainstorm and HTML prototypes before coding, interviews where Claude asks one question at a time, pointing Fable at reference source code, and an implementation-notes file to log deviations.
  • The author's own launch video for Fable was edited with Claude Code, using prototypes and a color-grading lesson to surface gaps in his knowledge.

The author's core claim is that working with Claude Fable 5 keeps bumping into one problem: the gap between what you've written down and what the actual work requires. They call the prompts, skills, and context "the map" and the codebase and real-world constraints "the territory." Every place where the map is silent is an "unknown," and when Claude hits one, it guesses at what you want. The more work in a task, the more guesses pile up. The author's sharpest point is that Fable is the first model where output quality is limited mainly by the user's ability to clarify those unknowns, not by the model itself. The framework is a four-part split. Known knowns are what you've already said. Known unknowns are gaps you're aware of. Unknown knowns are things so obvious you'd never write them down but would recognize on sight, like visual taste. Unknown unknowns are blind spots you haven't considered at all. The author argues the best agentic coders, pointing to Boris and Jarred as examples, have few unknowns to begin with, and that reducing and planning for them is the real skill. Being too specific makes Claude follow you past the point where a pivot would be better, while being too vague makes it fall back on generic best practices that may not fit. Giving Claude context about your starting point, your experience, and where you are in your thinking is presented as the main remedy. The article then lays out a set of practical techniques organized by phase. Before implementation, you run a "blind spot pass" asking Claude to surface your unknown unknowns, and you brainstorm or prototype cheaply, often as a single HTML mockup, because changing direction after code is written is expensive. You can also have Claude interview you one question at a time, prioritizing questions whose answers would change the architecture. When you can't describe what you want, point Claude at a reference, ideally source code in any language, since it can read the underlying implementation rather than just a screenshot. Finally, ask for an implementation plan that highlights the parts most likely to change, so you can review those before committing. The author notes they don't use every technique each time, and suggests the artifacts they built for this are worth returning to for building intuition about when each one fits.

Ask about this article