Skip to content

v0.1.2 — Anti-exhaustion release

Choose a tag to compare

View all tags
@nicobailon nicobailon released this 20 Jul 22:58
v0.1.2
b2ae474

Addresses the failure mode where "grilling" primes an agent to over-question to the point of user exhaustion, plus three refinements from Thariq's interview on how the Claude Code team plans with AI.

  • Defined grilling as few, evidence-priced questions; not asking about non-material topics is correct behavior.
  • Explicit stop condition: the grill ends when the unknowns ledger is empty, with remaining-count announcements.
  • Default budget of ~5 blocking questions per session; exceeding it requires user opt-in.
  • Fatigue valve: short/impatient answers switch the agent to defaults mode with one batched assumptions list for veto.
  • Batched endgame: residual low-risk unknowns presented as a single assumptions list instead of serial questions.
  • Bad-question guard: asking the user to verbalize taste they can only recognize when shown routes to prototypes/references instead.
  • Subjective verification: captured reactions distill into a short rubric that serves as the verification gate when no test can.
  • Blindspot pass now explicitly searches documented limits and known failure modes of load-bearing dependencies.

Full changelog: https://github.com/nicobailon/grill-for-unknowns/blob/main/CHANGELOG.md

Assets 2
Loading