Token Optimizer

Expert in maximizing LLM value while minimizing token consumption and API costs.

When to Use This Skill

• Reducing LLM API costs (OpenAI, Anthropic, Google)
• Optimizing prompts for token efficiency
• Implementing prompt caching strategies
• Managing context window limits
• Designing cost-aware AI architectures
• Compressing context without losing quality
• Building token usage monitoring

Core Workflow

1. Audit current usage - Measure tokens per request, identify waste
2. Compress prompts - Remove redundancy, use abbreviations, merge similar instructions
3. Optimize context - Summarize history, prioritize relevant information
4. Implement caching - Cache common prefixes, system prompts, few-shot examples
5. Monitor costs - Track token usage per feature, set budgets

Token Reduction Techniques

Prompt Compression

# Before (verbose - 150 tokens)
You are a helpful assistant that helps users with their coding questions.
When you receive a question, you should analyze it carefully and provide
a comprehensive answer. Make sure to include code examples when relevant.

# After (compressed - 45 tokens)
Role: Coding assistant. Analyze questions, provide answers with code examples.

System Prompt Optimization

# Before (redundant)
- Do not include unnecessary information
- Be concise
- Avoid repetition
- Keep answers short
- Don't ramble

# After (deduplicated)
Constraints: Be concise. No redundancy.

Few-Shot Example Optimization

# Before (full examples - 500 tokens)
Example 1: [100 tokens]
Example 2: [100 tokens]
Example 3: [100 tokens]

# After (compressed - 150 tokens)
Examples:
- Input: X -> Output: Y
- Input: A -> Output: B
- Input: C -> Output: C

Caching Strategies

Cost Comparison Guide

Monitoring Template

const trackTokens = (model, inputTokens, outputTokens) => {
  const cost = calculateCost(model, inputTokens, outputTokens);
  console.log(`Tokens: ${inputTokens} in / ${outputTokens} out / $${cost}`);
};

Constraints

MUST DO

• Measure token usage before and after optimization
• Maintain response quality while reducing tokens
• Implement streaming for long responses
• Cache stable prefixes (system prompts, few-shot)
• Use smaller models for simple tasks
• Set token budgets per feature

MUST NOT DO

• Sacrifice accuracy for token savings
• Skip quality validation after compression
• Hardcode token limits (make configurable)
• Ignore model-specific tokenization differences
• Cache sensitive/user-specific data

Knowledge Reference

Tokenization, tiktoken, prompt engineering, context windows, semantic caching, KV-cache, prompt compression, cost optimization, model routing, streaming responses