Claude Code's initial instructions to a model are a dump of 80kb of text. And none of it has anything to do with security. The "security" is handled by the eval

Claude Code's initial instructions to a model are a dump of 80kb of text. And none of it has anything to do with security. The "security" is handled by the evaluator model that sometimes denies some tool uses. I'm sure that in itself can be done very cheaply.