---

name: writing-rules description: "Use when creating or updating rules in CLAUDE.md, settings, or rule files. Covers confidence thresholds and false positive prevention."

Writing Rules

Overview

Rules are constraints that guide behavior. Good rules are specific, testable, and have clear thresholds for when they apply.

When to Create

Create a rule when:

• Same guidance given 3+ times manually
• Behavior needs to be consistent across sessions
• Mistake pattern keeps recurring
• Process discipline is required

Don't create for:

• Occasional edge cases (document as patterns instead)
• User preferences that change frequently
• One-time instructions

Structure Template

Copy## Rule Name

**When:** [Specific trigger conditions]
**Confidence:** [high/medium/low - when to apply vs ask]
**Action:** [What to do]

### Examples
- Good: [Correct application]
- Bad: [Incorrect application]

### Exceptions
[When this rule doesn't apply]

Quality Checklist

• Clear trigger condition (not vague)
• Confidence threshold defined
• Both good AND bad examples
• Exceptions explicitly listed
• Testable (can verify rule was followed)
• No overlap with existing rules
• Assigned to correct category (process/domain/project)

Testing Requirements

1. Create test cases - 5+ scenarios where rule should apply
2. Create negative cases - 5+ scenarios where rule should NOT apply
3. Run through agent - Verify correct activation
4. Measure false positive rate - Target < 10%
5. Monitor for 3 sessions - Auto-rollback if quality drops

Examples

Good Rule:

Copy## Import Path Verification

**When:** Adding or modifying import statements
**Confidence:** high
**Action:** Verify path exists by checking 3+ existing examples in codebase

### Examples
- Good: Check `src/components/Button.tsx` exists before importing
- Bad: Assume `@/components/Button` works without verification

### Exceptions
- Standard library imports (React, Node built-ins)
- Well-known packages from package.json

Bad Rule:

Copy## Be Careful

**When:** Doing things
**Action:** Think before acting

(Too vague, not testable, no examples)

Common Mistakes