context-engineering

To optimize AI agent context, structure information from persistent to transient using a hierarchical approach: Level 1 rules files like CLAUDE.md persist across sessions, while Level 2 and 3 specs and source files load per feature to prevent context flooding.

Context engineering is the structured loading of rules, specs, and source files into an AI agent's context stack. You need it during new session initialization, codebase area switching, and project onboarding to maximize output quality and prevent context drift.

Create a CLAUDE.md file encoding your project conventions and load it as persistent Level 1 context for all agents. This ensures rules persist across sessions to enforce boundaries and maintain consistent agent behavior.

Agent output drifts when context becomes flooded or leaks between features. Prevent this by applying a rules-based context stack that selectively loads per-feature specs and relevant source files only when needed, avoiding overwhelming the agent.

Yes, context engineering applies during project onboarding and when switching between codebase areas. It uses a hierarchical context stack to guide what information the agent sees, preventing context flooding in large codebases.