# Refine Rule

> Audit and slim always-loaded guidance: AGENTS.md, CLAUDE.md, Cursor rules.

- Skill: `mia-cx/refine-rule` (Agent Skill)
- Install (CLI): `npx skillmds@latest add mia-cx/refine-rule`
- Raw SKILL.md: https://api.skillmd.com/api/skills/mia-cx/refine-rule/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Security
- Author: mia-cx (https://skillmd.com/u/mia-cx)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/mia-cx/refine-rule

---


# Refine rule

Always-loaded files pay their cost every session; prune hardest here. Apply [agent-copy](../agent-copy/SKILL.md); Read it first if not already loaded. Checks specific to rule and memory files:

- **Right file for the audience.** Harness-agnostic guidance in AGENTS.md; harness-specific mechanics in that harness's file (CLAUDE.md, `.cursor/rules/*.mdc`). Cursor rules scope by glob; prefer a scoped rule over an always-apply one.
- **When a skill covers the topic**, the memory file keeps one pointer line. Restated bullets drift from the skill.
- **Delete what context will supply.** "When working on X, do Y" is suspect: Y is often self-evident inside X, visible in the files being edited or surfaced by typecheck, lint, or tests. *Can* be found ≠ *will* be encountered; facts in obscure corners stay written down.
- **Under 50 lines per section.** Longer is sediment.
- **Verify every fact.** Paths, model names, and tool flags rot silently in files nothing typechecks.