Practical Guide to Slimming Down the Context Window for AI Coding Agents

You are an expert in optimizing context window usage for AI coding agents. I need you to help me create a comprehensive context optimization strategy. ## My Setup - Coding agent: [Claude Code / Cursor / Copilot / Other] - Typical project size: [small/medium/large] - Main languages: [list languages] - Current pain points: [e.g., agent forgets context, slow responses, hitting token limits] ## Please provide: 1. **Context Audit**: Analyze what typically fills the context window in my workflow and identify waste 2. **File Strategy**: Which files should be in context vs. retrieved on-demand? Create a .contextignore or equivalent config 3. **Memory Architecture**: Design a tiered memory system: - Hot context (always loaded) - Warm context (loaded per-task) - Cold context (searchable but not loaded) 4. **Prompt Compression**: Rewrite my typical prompts to be 50%+ shorter while preserving intent 5. **Tool Output Sandboxing**: Strategy to prevent verbose tool outputs from consuming context 6. **Session Management**: When to start fresh vs. continue, and how to carry forward key decisions 7. **Metrics**: How to measure context efficiency (useful tokens / total tokens ratio) Output a concrete, copy-paste-ready configuration and workflow guide.