course-builderThis is an agent-run eval, not an automated grader. cases.yaml has three blocks:\nshould_trigger (7), should_not_trigger (5), and capability (1 scenario with a rubric). To run\nit, load only course-builder into a fresh agent (no sibling skills, so a miss can't be masked),\npaste each should_trigger / should_not_trigger prompt verbatim over 3–5 trials, and record\nwhether the agent invokes the skill — should_trigger should fire, should_not_trigger should\nstay quiet and ideally route to the named sibling. Aim for ≥90% correct decisions across trials; a\nprompt that misses on a majority of its trials is a real defect — fix the description/triggers in\nSKILL.md, don't loosen the case. For the capability scenario, run it once and score the output\nagainst the must_include rubric by hand: did it work backward (outcomes first), ban vanity verbs,\nverb-match assessment to outcomes with both formative and summative, sequence modules with no\norphans/unproven outcomes, emit the alignment matrix, and hand storytelling off rather than doing\nit? Grading is judgement, not grep — scripts/verify.sh only checks the greppable structure of a\nfinished curriculum doc (banlist verbs, matrix presence, proven outcomes, formative+summative); the\nrubric here is the real bar.