Give the model what it needs, not everything you have. 85% smaller reads · 97.4% image compression · 180+ filter & interception rules · 94–99% skill overhead cut · compaction memory · prompt injection ...