Create Rules and Skills

Rule Types (3-Tier System)

Type	Frontmatter	When loaded	Filename suffix
Always	alwaysApply: true, description/globs blank	Every conversation	-always.mdc
Auto	alwaysApply: false, globs filled	When matching file is open	-auto.mdc
Agent	alwaysApply: false, description filled	Agent decides	-agent.mdc
Manual	alwaysApply: false, both blank	Only via @ mention	-manual.mdc

Token Budget (CRITICAL)

• Total always-apply: TARGET < 5,000 tokens
• alwaysApply: true ONLY for rules needed in EVERY chat
• File-specific rules -> auto (globs)
• Task-specific rules -> agent (description) or skill
• Rule of Three: codify a rule ONLY after 3 repetitions of the same mistake

Naming Convention

{CATEGORY}-rule-name-{always|auto|agent|manual}.mdc

Categories: core, _base, protocol, standard, workflows, error

Rule Structure

See template in rule-template.mdc.

Description Format (Agent Rules)

ACTION-TRIGGER-OUTCOME format: "Применять при [TRIGGER]. [ACTION] для [OUTCOME]."

Universality Requirement

Rules must work in ANY project:

• NO project entities -> use placeholders
• NO project URLs/creds -> .cursor/.secrets/
• NO project architecture -> AGENTS.md
• YES: universal principles, abstract patterns

When to Create a Skill Instead of a Rule

• Workflow with multiple steps -> Skill
• Needs executable scripts -> Skill
• Long procedural content (>200 lines) -> Skill
• Only relevant for specific tasks -> Skill
• Short constraint needed broadly -> Rule

After Creating/Updating

1. Re-read the file (self-check)
2. Output: Rule/Skill path, Type, Description