context-engineering

To load project rules and align agent context, apply context engineering techniques by loading Level 1 rules files (like CLAUDE.md), per-feature specs, and relevant source files when starting a new session or switching tasks.

Context engineering is the practice of curating what information an agent loads across sessions and tasks. You need it when starting new projects or changing scope to reduce hallucinations and ensure the agent sees the correct project-wide rules.

To set up guardrails against data leakage, structure your context hierarchy from persistent rules to transient data, ensuring the agent loads only the correct project-wide rules, specs, and source files required for the current feature.

Yes, you can create and load rules files like CLAUDE.md to persistently govern agent behavior across sessions. This ensures the agent consistently follows project-wide conventions and architectural guidance.

The best way to structure context hierarchy is to organize information from persistent project rules to transient feature specs, loading relevant spec sections and source files only when needed to maintain focus and guard against misconfiguration.

Agents lose alignment when switching tasks because context is not properly refreshed. To fix this, apply context engineering to reload project rules, relevant spec sections, and feature-specific source files at the start of each new task.