Cursor Rule-of-Three: Codify .mdc Rules Only After Three Repeated AI Failures, Not One
A documented anti-pattern in Cursor .mdc rule engineering: developers add rules after the first AI mistake, causing rapid rule bloat that increases the 'token tax' (context overhead per request) without measurable accuracy gains. The Rule of Three pattern defers codification until the AI has violated the same architectural pattern three separate times — transforming anecdote into signal. The practical test: would a new team member violate this same rule without being told? If yes, add it; if a single code review comment would prevent it, don't. Hierarchically, rules should be organized as 00-project-context.mdc → 01-architecture.mdc → 99-agent-behavior.mdc, with earlier files taking precedence on conflict.
Source
↳ Follow the thread