Most people who struggle with AI agents have the same problem: they dump too much information into the prompt, or too little. Context engineering is the discipline of giving the agent exactly what it needs — no more, no less. Rule 1: Path-scope your rules Don't load API validation rules when the agent is editing CSS. Use path-scoped instruction files that only activate when the agent touches matching files. API rules for src/api/**/*.ts, component rules for components/**/*.tsx. Rule 2: Compact at transitions When you switch from backend work to frontend work, compress your conversation history. Stale context from a debugging session will confuse the agent when it's now building a UI component. Rule 3: Use @ imports sparingly Only attach files that are essential, short (<200 lines), and stable. For large documentation, create a skill file instead — skills load on demand, not on every turn. Rule 4: Keep your CLAUDE.md lean Every line in your instruction file costs ~4 tokens per turn. A 500-line CLAUDE.md burns 2,000 tokens on every single message. Keep it under 200 lines. Rule 5: Describe WHAT, not HOW Bad: "Please look at my code and find any bugs or performance issues. Make sure to check everything carefully." Good: "Audit src/api/ for N+1 queries and unhandled promise rejections." The expert prompt is bounded, concrete, and parallelizable. The beginner prompt is vague and forces the agent to guess. Teams that master context engineering see 3-5x better results from the same AI tools their competitors are using poorly.