A rule turns a decision into operating memory. It tells the next person or agent how to handle the work, so nobody has to rediscover the decision from scratch.
Think of the note taped above the front counter: “Do not send refunds without manager approval.” “Put every signed contract in this folder.” The note isn’t the whole business. It’s one decision made visible at the exact place people keep messing it up. A rule can look like a preference if you squint, but a preference (“I like this tone better”) needs interpretation every time, while a rule (“Don’t use this phrase in client-facing writing”) is specific enough that a future person or agent can follow it.
How it shows up
This matters with AI because the model doesn’t absorb your corrections forever. Fix an email by hand but never write down the pattern, and the next draft repeats the mistake. Tell the agent, “Notice what we changed, explain the rule, and save it,” and the correction becomes part of the system. Rules live in a few places. A skill can contain rules for a repeatable workflow. Project instructions hold rules for a project. A hook can enforce a rule automatically, like blocking an outbound email unless the right approval phrase exists. Guardrails are often rules with teeth, especially around safety or client trust.
Why you care
A good rule is narrow. “Write clearly” is too vague. “Use one sentence per line in email drafts” is a rule. If a rule can’t be followed or checked, it’s still too soft, because it should give the next session a clear behavior, not a mood. Rules and practice work together: the practice is the repeated way of working, and the rule keeps it from drifting when work gets busy or handed to an agent. Every useful AI system eventually becomes a memory problem, and rules are how you stop good judgment from disappearing after one session.