engineering-place-guardrails
Decide where a guardrail belongs so a mistake is prevented by the cheapest layer that can carry it: the codebase itself (API design, architecture, refactoring), then static analysis (types, compilers, lints, tests, CI), then verification (formal, property, runtime), then prose (rules, skills, AGENTS.md), and only last human review. Walks a proposed rule or a recurring mistake down that pyramid, stops at the first layer that can hold it, and reports what to build there plus what prose or review can be deleted once it lands. Use when someone wants to add a rule to AGENTS.md, CLAUDE.md, a skill, a style guide, or a review checklist; when the same review comment or agent mistake keeps coming back; when auditing an instructions file that keeps growing; when asking 'how do we stop agents from doing X', 'should this be a lint', or 'where should this convention live'. Do not use for choosing a specific lint tool's configuration, writing the test itself, or general code review.