What problem does it solve? AI coding agents can run dangerous commands or edit protected files in a repository. This Skill guides you in writing a project-level pi-gate configuration (.agents/anchors.json) so guardrails match the specific repo instead of relying only on global rules. ## Core Features & Use Cases - Layered authority model: Understand the three-layer ratchet (code floor, global config, project gate) where each layer can only add restrictions, never remove them. - Mechanism selection: Choose between frozen_commands (substring hard block), frozen_paths (prefix/suffix block), frozen_globs (pattern block), redirect_conventions (notify), rewrite (token swap), and path_hints (post-edit reminders). - Safety invariants: Respect meta-frozen global files, lock ratified project rules with _meta_frozen, and avoid substring pitfalls like single-word frozen commands. - Use Case: In a Terraform production repo, freeze "terraform apply" and "kubectl delete", protect the prod/ directory, and lock the gate with _meta_frozen so agents cannot weaken it. ## Quick Start Ask the agent to write a project gate for this repo that blocks deployments and protects the infra directory using .agents/anchors.json.
