I Let My AI Design Its Own Rules. Then It Broke Every Single One.

My AI assistant designed its own safeguard system. 500+ lines of rules. 9 custom hooks. Persistent memory files. A 258-file knowledge vault. Protocols it wrote, named, and documented. Then it violated every single one during a routine task. This is not a rant. This is an engineering post-mortem on AI self-governance. What I Built (With Claude's Help) I use Claude Code daily — Anthropic's CLI-based AI coding assistant. Over weeks of collaboration, I let Claude design and iterate on its own rule system. The stack: 1. CLAUDE.md — The Constitution (500+ lines) Claude Code reads a CLAUDE.md file at session start. Think of it as system instructions the AI loads before doing anything. Mine grew to 500+ lines, each rule born from a real failure: Date Failure Rule Created 2026-03-06 Proposed a solution without searching first, nearly wasted an hour "Search Before Speaking" iron rule 2026-03-07 Said "saved" twice when asked. Never wrote to disk. "ATOMIC SAVE PROTOCOL" 2026-03-08 258 knowledge fi