What problem does it solve? Prompts sent to expensive or autonomous AI agents often fail late and expensively because the task frame was never verified. This Skill defines normative sufficiency conditions for Initial Boundary Conditions (IBCs) and a cheap-tier adversarial refinement loop that catches bad frames before an expensive walk launches. ## Core Features & Use Cases - Seven Sufficiency Conditions (S1–S7): Checkable criteria covering falsifiable premises, rejection genres, decision rights (RESOLVED/DELEGATED/RESERVED), evaluator attachment, verbatim context curation, amendment protocols, and discipline proportion. - Boundary Refinement Loop (/boundary): A DRAFT → ATTACK → REVISE → APPROVE → DISPATCH cycle using isolated cheap-tier adversarial reviewers whose objections must cite a violated condition and evidence. - Comprehension Probe: A zero-context dry run where a cheap agent produces a plan from only the IBC, surfacing unanswered questions, plan-vs-intent divergence, and canary-trap bites before dispatch. - Use Case: Before dispatching an autonomous worker on a multi-hour coding campaign, run the refinement loop and probe to confirm the IBC's premises are verifiable, its decision rights are partitioned, and a human has approved the fixed-point IBC*. ## Quick Start Ask the agent to audit this task prompt against the boundary sufficiency conditions and run the refinement loop until it is ready to dispatch.