A Field Guide to Fable: Finding Your Unknowns

Takeaways

• Prompts/skills/context are the map; the codebase and real constraints are the territory—gaps between them are unknowns where the agent guesses.
• With Fable-class models, output quality is often bottlenecked by how well you surface and resolve unknowns, not raw model capability.
• Taxonomy: known knowns (in prompt), known unknowns (aware gaps), unknown knowns (tacit taste), unknown unknowns (blind spots).
• Unknowns appear before, during, and after implementation; upfront planning alone does not eliminate them.
• Techniques include blind-spot passes, HTML prototypes for taste, structured interviews, and post-implementation retros—pick by unknown type.
• Top agentic coders reduce unknowns by staying in sync with both codebase and model behavior; over-specifying vs. vagueness both fail when unknowns are unexamined.

Notes

Directly applicable to vault agent work: clarifying unknowns in routing rules, skill boundaries, and “what good looks like” is the leverage point for reliable /kb-triage, /clip, and coding sessions—not more prompt volume.

Open questions

• Which unknown-discovery patterns belong as first-class vault skills vs. one-off session prompts?