Workflow Optimization

Token Efficiency

Use Prompt Caching

Claude Code automatically implements prompt caching with 90% discount on cache reads:

• System prompt cached globally (first-party)
• Tool schemas cached per session
• Early conversation history cached after compression

Choose Right Model

• Sonnet: Fast, cost-effective for routine tasks
• Opus: Complex reasoning, architecture design
• Use /model command to switch

CLAUDE.md Best Practices

Minimalist Approach

Include only universal rules. Overly long CLAUDE.md consumes valuable context window space.

Hierarchical Organization

~/.claude/CLAUDE.md     → Cross-project preferences
project/CLAUDE.md       → Project architecture
.claude/rules/*.md      → Conditional rules
CLAUDE.local.md         → Local private (not committed)

Anti-Patterns

• ❌ Don't repeat what model already knows
• ❌ Don't add excessive formatting rules
• ❌ Don't place sensitive information (API keys)

Conversation Management

Use Plan Mode for Analysis

/plan

Read-only mode for safe code exploration before making changes.

Clear Context When Needed

/clear

Starts fresh conversation when context becomes too large or confused.

Use Worktree for Experiments

/worktree

Isolate experimental changes in Git worktree.

Tool Usage Tips

Be Specific with File Operations

# Good
Read src/components/Button.tsx lines 1-50

# Vague
Read the button file

Use Grep for Large Codebases

Search for all usages of useState

Leverage Skills

/commit
/review-pr
/verify

Cost Management

Monitor Token Usage

/usage

Check current session token consumption.

Use Headless Mode for Scripts

claude -p "Generate a README for this project"

No TUI overhead, direct stdout output.

Enterprise Tips

Configure Policy Limits

• Set tool restrictions via enterprise policy
• Block specific MCP servers if needed
• Enable audit logging for compliance

Team Memory Sync

• Share knowledge across team members
• Secret scanning enabled by default
• Server-preferred sync semantics