rule-curator

原文涵盖的主题

• Rule Curator
• Overview
• When to use
• Workflow
• 1. Discover
• 2. Scope gate (required, defaults to proceed)
• 3. Extract
• 4. Analyze (against the whole corpus)

Agents accumulate behavioral rules across many files. Nobody prunes them, so they go stale, duplicate each other, duplicate installed skills, restate default behavior, or impose rituals whose friction exceeds their value. Core principle: every rule is a permanent tax. It costs context tokens every session and dilutes instruction-following attention across more directives. A rule earns its place only if its behavioral payoff beats that ongoing cost. This skill audits the rule set against that bar and acts on the result. This skill exists because a capable model, left to its own devices, finds the loud defects (conflicts, staleness, duplication) but skips the quiet ones (brittle rules, misfiled entries), never separates a rule's neutral effect from the critique of it, and produces prose instead of a usable deliverable. The workflow below forces all four. "Audit / clean up / prune my CLAUDE.md / rules / memory" After a big refactor or decision, to find rules that went stale The rule set has grown and instruction-following feels diluted Periodic hygiene on the directives an agent operates under Not for: authoring new rules (this curates existing ones), or auditing PRDs and feature spec

这里只展示节选，请以完整原文和当前要求为准。