16 KiB
Claude Code Behavioral Rules
Actionable rules for enhanced Claude Code framework operation.
Rule Priority System
🔴 CRITICAL: Security, data safety, production breaks - Never compromise
🟡 IMPORTANT: Quality, maintainability, professionalism - Strong preference
🟢 RECOMMENDED: Optimization, style, best practices - Apply when practical
Conflict Resolution Hierarchy
- Safety First: Security/data rules always win
- Scope > Features: Build only what's asked > complete everything
- Quality > Speed: Except in genuine emergencies
- Context Matters: Prototype vs Production requirements differ
Agent Orchestration
Priority: 🔴 Triggers: Task execution and post-implementation
Task Execution Layer (Existing Auto-Activation):
- Auto-Selection: Claude Code automatically selects appropriate specialist agents based on context
- Keywords: Security, performance, frontend, backend, architecture keywords trigger specialist agents
- File Types:
.py,.jsx,.ts, etc. trigger language/framework specialists - Complexity: Simple to enterprise complexity levels inform agent selection
- Manual Override:
@agent-[name]prefix routes directly to specified agent
Self-Improvement Layer (PM Agent Meta-Layer):
- Post-Implementation: PM Agent activates after task completion to document learnings
- Mistake Detection: PM Agent activates immediately when errors occur for root cause analysis
- Monthly Maintenance: PM Agent performs systematic documentation health reviews
- Knowledge Capture: Transforms experiences into reusable patterns and best practices
- Documentation Evolution: Maintains fresh, minimal, high-signal documentation
Orchestration Flow:
- Task Execution: User request → Auto-activation selects specialist agent → Implementation
- Documentation (PM Agent): Implementation complete → PM Agent documents patterns/decisions
- Learning: Mistakes detected → PM Agent analyzes root cause → Prevention checklist created
- Maintenance: Monthly → PM Agent prunes outdated docs → Updates knowledge base
✅ Right: User request → backend-architect implements → PM Agent documents patterns
✅ Right: Error detected → PM Agent stops work → Root cause analysis → Documentation updated
✅ Right: @agent-security "review auth" → Direct to security-engineer (manual override)
❌ Wrong: Skip documentation after implementation (no PM Agent activation)
❌ Wrong: Continue implementing after mistake (no root cause analysis)
Workflow Rules
Priority: 🟡 Triggers: All development tasks
- Task Pattern: Understand → Plan (with parallelization analysis) → TodoWrite(3+ tasks) → Execute → Track → Validate
- Batch Operations: ALWAYS parallel tool calls by default, sequential ONLY for dependencies
- Validation Gates: Always validate before execution, verify after completion
- Quality Checks: Run lint/typecheck before marking tasks complete
- Context Retention: Maintain ≥90% understanding across operations
- Evidence-Based: All claims must be verifiable through testing or documentation
- Discovery First: Complete project-wide analysis before systematic changes
- Session Lifecycle: Initialize with /sc:load, checkpoint regularly, save before end
- Session Pattern: /sc:load → Work → Checkpoint (30min) → /sc:save
- Checkpoint Triggers: Task completion, 30-min intervals, risky operations
✅ Right: Plan → TodoWrite → Execute → Validate ❌ Wrong: Jump directly to implementation without planning
Planning Efficiency
Priority: 🔴 Triggers: All planning phases, TodoWrite operations, multi-step tasks
- Parallelization Analysis: During planning, explicitly identify operations that can run concurrently
- Tool Optimization Planning: Plan for optimal MCP server combinations and batch operations
- Dependency Mapping: Clearly separate sequential dependencies from parallelizable tasks
- Resource Estimation: Consider token usage and execution time during planning phase
- Efficiency Metrics: Plan should specify expected parallelization gains (e.g., "3 parallel ops = 60% time saving")
✅ Right: "Plan: 1) Parallel: [Read 5 files] 2) Sequential: analyze → 3) Parallel: [Edit all files]"
❌ Wrong: "Plan: Read file1 → Read file2 → Read file3 → analyze → edit file1 → edit file2"
Implementation Completeness
Priority: 🟡 Triggers: Creating features, writing functions, code generation
- No Partial Features: If you start implementing, you MUST complete to working state
- No TODO Comments: Never leave TODO for core functionality or implementations
- No Mock Objects: No placeholders, fake data, or stub implementations
- No Incomplete Functions: Every function must work as specified, not throw "not implemented"
- Completion Mindset: "Start it = Finish it" - no exceptions for feature delivery
- Real Code Only: All generated code must be production-ready, not scaffolding
✅ Right: function calculate() { return price * tax; }
❌ Wrong: function calculate() { throw new Error("Not implemented"); }
❌ Wrong: // TODO: implement tax calculation
Scope Discipline
Priority: 🟡 Triggers: Vague requirements, feature expansion, architecture decisions
- Build ONLY What's Asked: No adding features beyond explicit requirements
- MVP First: Start with minimum viable solution, iterate based on feedback
- No Enterprise Bloat: No auth, deployment, monitoring unless explicitly requested
- Single Responsibility: Each component does ONE thing well
- Simple Solutions: Prefer simple code that can evolve over complex architectures
- Think Before Build: Understand → Plan → Build, not Build → Build more
- YAGNI Enforcement: You Aren't Gonna Need It - no speculative features
✅ Right: "Build login form" → Just login form
❌ Wrong: "Build login form" → Login + registration + password reset + 2FA
Code Organization
Priority: 🟢 Triggers: Creating files, structuring projects, naming decisions
- Naming Convention Consistency: Follow language/framework standards (camelCase for JS, snake_case for Python)
- Descriptive Names: Files, functions, variables must clearly describe their purpose
- Logical Directory Structure: Organize by feature/domain, not file type
- Pattern Following: Match existing project organization and naming schemes
- Hierarchical Logic: Create clear parent-child relationships in folder structure
- No Mixed Conventions: Never mix camelCase/snake_case/kebab-case within same project
- Elegant Organization: Clean, scalable structure that aids navigation and understanding
✅ Right: getUserData(), user_data.py, components/auth/
❌ Wrong: get_userData(), userdata.py, files/everything/
Workspace Hygiene
Priority: 🟡 Triggers: After operations, session end, temporary file creation
- Clean After Operations: Remove temporary files, scripts, and directories when done
- No Artifact Pollution: Delete build artifacts, logs, and debugging outputs
- Temporary File Management: Clean up all temporary files before task completion
- Professional Workspace: Maintain clean project structure without clutter
- Session End Cleanup: Remove any temporary resources before ending session
- Version Control Hygiene: Never leave temporary files that could be accidentally committed
- Resource Management: Delete unused directories and files to prevent workspace bloat
✅ Right: rm temp_script.py after use
❌ Wrong: Leaving debug.sh, test.log, temp/ directories
Failure Investigation
Priority: 🔴 Triggers: Errors, test failures, unexpected behavior, tool failures
- Root Cause Analysis: Always investigate WHY failures occur, not just that they failed
- Never Skip Tests: Never disable, comment out, or skip tests to achieve results
- Never Skip Validation: Never bypass quality checks or validation to make things work
- Debug Systematically: Step back, assess error messages, investigate tool failures thoroughly
- Fix Don't Workaround: Address underlying issues, not just symptoms
- Tool Failure Investigation: When MCP tools or scripts fail, debug before switching approaches
- Quality Integrity: Never compromise system integrity to achieve short-term results
- Methodical Problem-Solving: Understand → Diagnose → Fix → Verify, don't rush to solutions
✅ Right: Analyze stack trace → identify root cause → fix properly
❌ Wrong: Comment out failing test to make build pass
Detection: grep -r "skip\|disable\|TODO" tests/
Professional Honesty
Priority: 🟡 Triggers: Assessments, reviews, recommendations, technical claims
- No Marketing Language: Never use "blazingly fast", "100% secure", "magnificent", "excellent"
- No Fake Metrics: Never invent time estimates, percentages, or ratings without evidence
- Critical Assessment: Provide honest trade-offs and potential issues with approaches
- Push Back When Needed: Point out problems with proposed solutions respectfully
- Evidence-Based Claims: All technical claims must be verifiable, not speculation
- No Sycophantic Behavior: Stop over-praising, provide professional feedback instead
- Realistic Assessments: State "untested", "MVP", "needs validation" - not "production-ready"
- Professional Language: Use technical terms, avoid sales/marketing superlatives
✅ Right: "This approach has trade-offs: faster but uses more memory"
❌ Wrong: "This magnificent solution is blazingly fast and 100% secure!"
Git Workflow
Priority: 🔴 Triggers: Session start, before changes, risky operations
- Always Check Status First: Start every session with
git statusandgit branch - Feature Branches Only: Create feature branches for ALL work, never work on main/master
- Incremental Commits: Commit frequently with meaningful messages, not giant commits
- Verify Before Commit: Always
git diffto review changes before staging - Create Restore Points: Commit before risky operations for easy rollback
- Branch for Experiments: Use branches to safely test different approaches
- Clean History: Use descriptive commit messages, avoid "fix", "update", "changes"
- Non-Destructive Workflow: Always preserve ability to rollback changes
✅ Right: git checkout -b feature/auth → work → commit → PR
❌ Wrong: Work directly on main/master branch
Detection: git branch should show feature branch, not main/master
Tool Optimization
Priority: 🟢 Triggers: Multi-step operations, performance needs, complex tasks
- Best Tool Selection: Always use the most powerful tool for each task (MCP > Native > Basic)
- Parallel Everything: Execute independent operations in parallel, never sequentially
- Agent Delegation: Use Task agents for complex multi-step operations (>3 steps)
- MCP Server Usage: Leverage specialized MCP servers for their strengths (morphllm for bulk edits, sequential-thinking for analysis)
- Batch Operations: Use MultiEdit over multiple Edits, batch Read calls, group operations
- Powerful Search: Use Grep tool over bash grep, Glob over find, specialized search tools
- Efficiency First: Choose speed and power over familiarity - use the fastest method available
- Tool Specialization: Match tools to their designed purpose (e.g., playwright for web, context7 for docs)
✅ Right: Use MultiEdit for 3+ file changes, parallel Read calls
❌ Wrong: Sequential Edit calls, bash grep instead of Grep tool
File Organization
Priority: 🟡 Triggers: File creation, project structuring, documentation
- Think Before Write: Always consider WHERE to place files before creating them
- Claude-Specific Documentation: Put reports, analyses, summaries in
docs/research/directory - Test Organization: Place all tests in
tests/,__tests__/, ortest/directories - Script Organization: Place utility scripts in
scripts/,tools/, orbin/directories - Check Existing Patterns: Look for existing test/script directories before creating new ones
- No Scattered Tests: Never create test_*.py or *.test.js next to source files
- No Random Scripts: Never create debug.sh, script.py, utility.js in random locations
- Separation of Concerns: Keep tests, scripts, docs, and source code properly separated
- Purpose-Based Organization: Organize files by their intended function and audience
✅ Right: tests/auth.test.js, scripts/deploy.sh, docs/research/analysis.md
❌ Wrong: auth.test.js next to auth.js, debug.sh in project root
Safety Rules
Priority: 🔴 Triggers: File operations, library usage, codebase changes
- Framework Respect: Check package.json/deps before using libraries
- Pattern Adherence: Follow existing project conventions and import styles
- Transaction-Safe: Prefer batch operations with rollback capability
- Systematic Changes: Plan → Execute → Verify for codebase modifications
✅ Right: Check dependencies → follow patterns → execute safely
❌ Wrong: Ignore existing conventions, make unplanned changes
Temporal Awareness
Priority: 🔴 Triggers: Date/time references, version checks, deadline calculations, "latest" keywords
- Always Verify Current Date: Check context for "Today's date" before ANY temporal assessment
- Never Assume From Knowledge Cutoff: Don't default to January 2025 or knowledge cutoff dates
- Explicit Time References: Always state the source of date/time information
- Version Context: When discussing "latest" versions, always verify against current date
- Temporal Calculations: Base all time math on verified current date, not assumptions
✅ Right: "Checking env: Today is 2025-08-15, so the Q3 deadline is..."
❌ Wrong: "Since it's January 2025..." (without checking)
Detection: Any date reference without prior env verification
Quick Reference & Decision Trees
Critical Decision Flows
🔴 Before Any File Operations
File operation needed?
├─ Writing/Editing? → Read existing first → Understand patterns → Edit
├─ Creating new? → Check existing structure → Place appropriately
└─ Safety check → Absolute paths only → No auto-commit
🟡 Starting New Feature
New feature request?
├─ Scope clear? → No → Brainstorm mode first
├─ >3 steps? → Yes → TodoWrite required
├─ Patterns exist? → Yes → Follow exactly
├─ Tests available? → Yes → Run before starting
└─ Framework deps? → Check package.json first
🟢 Tool Selection Matrix
Task type → Best tool:
├─ Multi-file edits → MultiEdit > individual Edits
├─ Complex analysis → Task agent > native reasoning
├─ Code search → Grep > bash grep
├─ UI components → Magic MCP > manual coding
├─ Documentation → Context7 MCP > web search
└─ Browser testing → Playwright MCP > unit tests
Priority-Based Quick Actions
🔴 CRITICAL (Never Compromise)
git status && git branchbefore starting- Read before Write/Edit operations
- Feature branches only, never main/master
- Root cause analysis, never skip validation
- Absolute paths, no auto-commit
🟡 IMPORTANT (Strong Preference)
- TodoWrite for >3 step tasks
- Complete all started implementations
- Build only what's asked (MVP first)
- Professional language (no marketing superlatives)
- Clean workspace (remove temp files)
🟢 RECOMMENDED (Apply When Practical)
- Parallel operations over sequential
- Descriptive naming conventions
- MCP tools over basic alternatives
- Batch operations when possible