# Architecture, not prose **The rule.** A rule that lives only in prose is worth whatever the agent that happened to read it is worth. The same sentence as a hook, lint, script or test is worth the same on every run. So when a rule gets broken repeatedly, the fix is a mechanism, not a firmer sentence. ![When a rule keeps breaking, move it up: architecture, then compiler and tests, then CI and lint, then bot rules, then prose](../../assets/rule-ladder.png) The ladder is poteto's hierarchy: architecture makes the wrong thing impossible, the compiler and tests catch it before it runs, CI and lint refuse it, bot rules (hooks, skills) steer the agent, and prose is a note the agent may skip. A rule moves up a rung each time it breaks where it is. **Enforced by** the coverage tool in this directory, which maps every rule to its enforcer or marks it prose-only, and ranks the prose-only ones by how often they have actually been broken. **A miss the owner caught means a layer is missing** (the laptop, 11-15). When the person finds what the harness should have, the fix moves the rule up a rung, and the report says which layer enforces it now. A memory note, an instruction-file line, or "handed to another session" is still prose: it is a record of the miss, not a fix for it. Read that ranking as a queue of candidates, not a backlog. On the setup it was built for, about two thirds of the rules are still prose, and that is fine: see [guards need a real failure](guards-need-a-real-failure.md).