grillinglisted

# Attack the premise before it becomes code A wrong fix costs one cycle. A wrong frame costs every cycle until someone questions it, and nothing in the normal workflow questions it: review checks the diff against the plan, tests check the code against the spec, and both pass happily while the spec answers the wrong question. Adapted from Matt Pocock's `grilling` skill (github.com/mattpocock/skills, MIT). The tree-and-frontier mechanic is his. What is added here is the fact-finding rule made mandatory, the interaction shape this harness actually has, and an explicit list of what not to grill. ## The tree and the frontier Map the plan as a tree of decisions. Each decision branches into the decisions that hang off it: choosing a queue implies choosing a delivery guarantee, which implies choosing what a duplicate costs. The **frontier** is every decision whose prerequisites are already settled — the questions that can be answered right now, without waiting on another open one. Work in rounds: 1. Compute the frontier. 2. Ask the whole frontier in ONE round. Not one question at a time: a plan interrogated serially takes twenty turns and the user stops answering. 3. Each answer settles a decision and expands the frontier with whatever it unblocked. 4. Repeat until the frontier is empty. A question whose answer depends on an unresolved decision belongs to a later round. Asking it early produces an answer the user will retract. ## Finding facts is YOUR job **Never ask t