Rules Management

Rules Management Skill

Comprehensive guidance for managing CLAUDE.md user-specific rules with context-based loading for optimal performance.

Core Principle: Load What's Relevant

Claude Code loads CLAUDE.md on every interaction. To avoid loading all rules upfront:

1. Critical rules → Keep directly in CLAUDE.md (always loaded)
2. Context-specific rules → Move to .claude/rules/ with trigger hints (loaded on-demand)

Directory Structure

.claude/
├── CLAUDE.md                    # Critical rules + context hints
└── rules/                       # Domain-specific rules (loaded when relevant)
    ├── python.md
    ├── typescript.md
    ├── docker.md
    ├── react.md
    └── security-advanced.md

CLAUDE.md Structure (Always Loaded)

# Project Name

Brief project description.

## Critical Rules (Always Apply)

### Security
- Never commit secrets, API keys, or credentials
- Never disable security middleware or authentication checks
- Sanitize all user input before database operations

### Git Workflow
- Feature branches only, never commit directly to main
- Always run tests before committing

## Context (Read When Relevant)

Load context files based on what you're working on:

| Working On | Read |
|------------|------|
| Python/FastAPI backend | `.claude/rules/python.md` |
| React/TypeScript frontend | `.claude/rules/typescript.md` |
| Docker/containers | `.claude/rules/docker.md` |
| Authentication/security | `.claude/rules/security.md` |

What Goes Where?

Critical Rules (in CLAUDE.md)

Keep rules that are always relevant regardless of task:

• Security fundamentals (secrets, auth, input sanitization)
• Git workflow (branching, commits)
• Core project conventions (naming, structure)
• Communication preferences

Characteristics:

• Short and actionable
• Apply to 80%+ of tasks
• Safety-critical
• Never need context to understand

Context Files (in .claude/rules/)

Move rules that are domain-specific:

• Language-specific patterns (Python, TypeScript, Go)
• Framework conventions (React, FastAPI, Next.js)
• Tool configurations (Docker, Terraform, CI/CD)
• Detailed security practices

Characteristics:

• Only relevant for specific file types/tasks
• May be lengthy with examples
• Require specialized knowledge
• Not needed for every interaction

Context File Template

# [Technology/Domain] Context

Rules for [technology] development in this project.

## Key Patterns

- [Specific pattern or convention]
- [Another pattern]

## Examples

### Recommended
​```python
# Good example with explanation
​```

### Avoid
​```python
# Bad example with explanation why
​```

## Integration Points

- Related to: `.claude/context/[related].md` (when working with both)

Trigger Hints

The CLAUDE.md table tells Claude when to load context files from .claude/rules/:

Migration Guide

Before (All Loaded)

# CLAUDE.md

@rules/frontend/react.md
@rules/frontend/typescript.md
@rules/backend/python.md
@rules/backend/fastapi.md
@rules/devops/docker.md
@rules/security/auth.md
@rules/general/standards.md

Problem: All 7 files loaded on every interaction

After (Verrueckt Rules)

# CLAUDE.md

## Critical Rules
- Never commit secrets
- Feature branches only
- Run tests before commit

## Context (Read When Relevant)
| Working On | Read |
|------------|------|
| Python backend | `.claude/rules/python.md` |
| TypeScript/React | `.claude/rules/typescript.md` |
| Containers | `.claude/rules/docker.md` |

Improvement: Only load what's needed per task

Rule Priority

1. CLAUDE.md critical rules → Always apply
2. Loaded context files → Apply when read
3. Default best practices → Fall back when no rule exists

Best Practices

For Critical Rules

• Keep under 20 lines total
• Use bullet points, not paragraphs
• Focus on "never do" and "always do"
• No examples needed (rules should be self-evident)

For Context Files

• One technology/domain per file
• Include examples for complex patterns
• Reference related context files
• Keep under 200 lines per file

For Trigger Hints

• Use file extensions as triggers (.py, .tsx)
• Use technology keywords (Python, React)
• Use task keywords (auth, deploy, test)
• Be specific enough to avoid false matches

Commands Reference