LLM Token Optimization Rules

Overview​

Core Principles​

How LLM Pricing Works​

• Token-based billing: Cost is calculated per token, not per message
• Input + Output tokens: You pay for both what you send and what you receive
• Stateless nature: LLMs don't remember previous conversations
• Context resending: Every follow-up message includes the entire conversation history
• Compounding costs: Token usage grows exponentially with conversation length

Token Usage Examples​

Message 1: 1,000 input + 2,000 output = 3,000 tokens
Message 2: 4,000 input (includes Message 1) + 2,000 output = 6,000 tokens
Message 3: 7,000 input (includes Messages 1-2) + 2,000 output = 9,000 tokens

Optimization Rules​

Rule 1: Start New Chats for New Tasks​

• Use /clear command in Claude Code to wipe conversation history
• One chat window per task
• Start fresh when switching contexts
• Never reuse long chat threads for different tasks
• Don't let conversations grow beyond necessary scope

• When next task builds directly on the previous one
• Working on complex, interconnected features requiring ongoing context

Rule 2: Summarize Long Conversations​

• Use /compact command in Claude Code
• Provide custom summary instructions
• Focus summaries on relevant information only
• Trim unnecessary context and discussions

• "Please only summarize and keep the last message"
• "Focus on technical details, skip general discussion"
• "Summarize only action items and open issues"
• "Format summary as XML/JSON for structured data"

Rule 3: Choose the Right Model​

• Use /model command to switch models manually
• Start with powerful models for planning and architecture
• Switch to lighter models for implementation and refinement
• Don't default to most expensive model for simple tasks

Use Powerful Models (Claude Opus, GPT-4) For:​

• Complex reasoning and problem-solving
• System architecture design
• High-level planning and strategy
• Edge case identification
• Critical decision-making

Use Lighter Models (Claude Sonnet, GPT-3.5) For:​

• Code refactoring and cleanup
• Documentation writing
• Simple implementations
• Routine coding tasks
• Follow-up questions and clarifications

• Claude Opus: $15/1M input tokens, $75/1M output tokens
• Claude Sonnet: $3/1M input tokens, $15/1M output tokens

Best Practices​

Conversation Management​

1. Monitor context usage: Watch for capacity indicators (aim for <50%)
2. Be specific in requests: Avoid vague prompts that generate unnecessary output
3. Use focused follow-ups: Ask targeted questions rather than broad requests
4. Clean up regularly: Don't let conversations accumulate unnecessary history

Cost-Effective Workflows​

1. Planning Phase: Use powerful model for architecture and planning
2. Implementation Phase: Switch to lighter model for coding
3. Review Phase: Use appropriate model based on complexity of review needed
4. Documentation Phase: Use lighter model for writing and formatting

Token Waste Prevention​

• Avoid uploading irrelevant files or large documents
• Don't include unnecessary context in prompts
• Minimize repetitive back-and-forth on simple topics
• Don't let auto-selections choose overpowered models
• Be precise and direct in your requests
• Structure complex requests clearly
• Use appropriate formatting to reduce ambiguity

Advanced Techniques​

Custom Model Configuration​

/model claude-3-5-sonnet-20241022

• Copy exact model names from provider documentation
• Save frequently used models for quick switching
• Create model profiles for different types of work

Context Optimizations​

• Preemptive summarization: Summarize before hitting limits
• Strategic context: Only include relevant previous messages
• Structured prompts: Use clear formatting to reduce interpretation overhead
• Batch related questions: Group similar requests to minimize context switching

Performance vs. Cost Balance​

• Long-term projects: Invest in planning with powerful models upfront
• Iterative work: Use lighter models for incremental changes
• Critical tasks: Don't compromise on model quality for important decisions
• Routine tasks: Optimize aggressively for cost on repetitive work

Monitoring and Measurement​

• Track token usage patterns over time
• Monitor cost per task type
• Identify inefficient conversation patterns
• Measure model performance vs. cost for different task types
• Set budget alerts and usage limits

Quick Reference Commands​

Claude Code Commands​

• /clear - Start new conversation
• /compact - Summarize and compress conversation
• /model - Switch between models
• /model default - Return to default model selection
• /model claude-3-5-sonnet-20241022 - Use specific model

Decision Matrix​