What problem does it solve? When you dispatch an agent with a brief that prescribes a specific fix, threshold, or value, the agent may discover at the worksite that following your instruction literally makes things worse. A compliant agent ships the defect silently; an argumentative agent refuses without proof. This Skill defines a two-part protocol — explicit permission to satisfy the intent of an instruction while refusing its letter, plus a mandatory evidence requirement — so refusals become checkable facts instead of conversations. ## Core Features & Use Cases - Permission paragraph for briefs: A copy-paste block that authorizes deviation only when literal compliance defeats the intended outcome, and names evidence as the price of that permission. - Outcome-shaped instruction guidance: A method-to-outcome rewrite table (e.g., "set the timeout to 30s" becomes "the step must not be killed by its own timeout under observed load") that prevents most bad literal reads before dispatch. - Refusal review rubric: Three ordered checks — did the agent build or measure the thing, is the counterfactual stated, is the artefact clean (reverted unpushed) — for deciding whether to bank a reported deviation. - Use Case: An agent told to apply a pipeline fix builds it, measures that it converts a silent safe refusal into a permanently red pipeline, reverts the change unpushed, and reports the measurement — turning a potential defect into a verified fact. ## Quick Start Add the permission paragraph from this Skill to your next agent brief, requiring that any deviation from your instructions include what was asked, what was done instead, the measured evidence, and the counterfactual.
