A real guardrail is a hook in the harness, not an instruction in the prompt: the Claude Code hook mechanism, the config layers, then each AgentKit guard layer and the gaps worth checking.
Part 1/2: a prompt instruction is only a suggestion. A real guardrail is a hook in the Claude Code harness that blocks a tool call with exit 2. The hook mechanism, AgentKit's seven guardrail groups and its settings layers.
Part 2/2: each AgentKit guard layer taken apart, from scout-block, privacy-block and simplify-gate to hard-gates and guard skills, with the known gaps, habits worth keeping and a quick glossary.
