token-kill

# Token Kill - OpenClaw Token Optimizer Need help optimizing your OpenClaw token usage costs? This Skill will guide you through three powerful optimization techniques to dramatically reduce token consumption. Based on real-world case studies, applying these optimization techniques can reduce token consumption from **$200+/day** to **$10/day**, achieving a **95%+ cost reduction**. ## Three Core Token Optimization Techniques ### 1️⃣ Slash Commands Optimization - `/new` - Start a fresh conversation and clear old context (saves 50,000+ tokens) - `/compress` - Compress memory by keeping important info and forgetting details (saves 30,000+ tokens) - `/stop` - Immediately stop current task to prevent further token consumption - `/restart` - Restart the system to clear lag and resolve issues ### 2️⃣ Script-First Principle **Core Philosophy: AI is your brain, not your hands** Automate with scripts instead of using the model for mechanical tasks: - 📧 **Email Checking** - Scripts monitor emails; AI only notified of new messages ($100+/month → <$1/month) - 🌤️ **Weather Queries** - Direct API calls, zero token consumption - 📊 **Data Fetching** - Scripts retrieve data; AI only handles formatting - ⏰ **Scheduled Tasks** - Scripts execute; prevent AI from polling - 🔄 **Data Processing** - Script handles transformations ### 3️⃣ Model Tiering Strategy **Use premium models for complex tasks, budget models for simple ones** | Complexity | Recommended Model | Cost | Use Cases | Savings | |-----------|------------------|------|-----------|---------| | 🔴 High | GPT-4 / Claude | $0.03/1k tokens | Code generation, creative writing, complex reasoning | Baseline | | 🟡 Medium | GPT-3.5-Turbo / Ernie | $0.0005/1k tokens | General tasks, text editing | 98% | | 🟢 Low | Qwen, Tongyi (Budget Models) | $0.00001/1k tokens | Data processing, report generation, formatting | **99.97%** | ## Real-World Cost Reduction Cases ### Case 1: Email Monitoring System **Problem**: Model checks emails every 5 minutes | Approach | Monthly Cost | |----------|-------------| | ❌ Model Polling | **$100+/month** | | ✅ Script + AI Notification | **<$1/month** | | **Savings** | **99%** | ### Case 2: Daily Report Generation **Scenario**: Generate reports every 30 minutes (2000 tokens/call) | Model | Daily Cost | Monthly Cost | Savings | |-------|-----------|-------------|---------| | GPT-4 | $2.88 | $86 | Baseline | | GPT-3.5 | $0.048 | $1.44 | 98% | | Qwen | $0.001 | $0.03 | **99.97%** | ## Examples ### Example 1: Compressing Large Memory **Scenario**: After many conversations, memory.md has grown to hundreds of thousands of characters **Solution**: 1. Execute `/compress` command 2. System removes trivial details while preserving core information 3. Memory size reduced by 30-50% **Result**: Reduced context loading on each turn, saves 30,000+ tokens ### Example 2: Replacing AI with Scripts **Scenario**: Need to check for new orders every hour **Wrong Approach**: ``` Have model check orders API every hour → Model must understand and judge each time → 24 checks per day = huge costs ``` **Correct Approach**: ``` Script checks order API every hour Notify model only on new orders Model handles decision-making only ``` **Savings**: Script uses only CPU, saves 90%+ tokens ### Example 3: Model Tiering Workflow **Scenario**: Handle various complexity levels **Strategy**: - 💻 **Code Writing** → GPT-4 (worth the investment) - 📝 **Content Editing** → GPT-3.5 (good balance) - 📊 **Report Generation** → Budget Model (fully sufficient) **Result**: 90% cost reduction, zero functionality loss ## Guidelines ### ✅ Best Practices for Token Savings #### 1. Use Slash Commands Regularly - **Execute `/compress` once daily** - Prevent memory bloat - **Use `/new` for long conversations** - Start fresh after 1+ hours - **Use `/stop` on wrong tasks** - Stop immediately to prevent waste #### 2. Strictly Follow Script-First Principle - ✅ **Scripts handle**: Scheduled checks, data fetching, API calls, data processing - ❌ **Never let AI handle**: Polling, mechanical work, repetitive checks, resource-intensive operations - 💡 **Core rule**: AI = decision-making and judgment; Scripts = execution and heavy lifting #### 3. Enforce Model Tiering | Task Type | Model Choice | Reason | |-----------|-------------|--------| | Code generation, deep analysis | GPT-4 | Complex tasks worth the cost | | General tasks, text editing | GPT-3.5 | Best value proposition | | Data processing, reports | Budget Models | Fully capable, lowest cost | #### 4. Regular Token Usage Audit - Review billing distribution - Identify high-cost tasks for optimization - Adjust model configuration and scripts ### ❌ Common Token Wastage Patterns | Bad Practice | Consequence | Solution | |-------------|-------------|----------| | Unlimited conversation history | Growing memory = more tokens | Regular `/compress` or `/new` | | AI polling for updates | Token burn on each check | Use scripts instead | | Using GPT-4 for simple tasks | Overkill, high cost | Use appropriate model tier | | Never compressing memory | Linear token cost growth | Establish compression habit | | Continuing failed tasks | Wasted tokens | Use `/stop` immediately | ## Token Cost Formula ``` Total Cost = Context Consumption + Task Consumption Optimization Formula: New Cost = (Original Context × 30%) + (Task Cost × 20%) = Original Cost × (0.3 + 0.2) = Original Cost × 0.5 or lower ``` Combining all three techniques achieves **95%+ cost reduction**. ## Key Principle > 💡 **Remember**: High costs don't come from AI itself, but from making it do tasks it shouldn't do and remember information it shouldn't store. > > Assign the right tasks to the right tools, and AI becomes truly cost-effective.