context-engineering

Manage long agent conversations by summarizing progress and compacting context when the session grows. This hierarchical context model maintains stability and ensures the agent retains critical task information without exceeding limits.

Context engineering works by selectively loading relevant spec sections and necessary source files before editing. This prevents loading entire documents and ensures the agent only sees task-relevant information, reducing hallucinations.

Guide agent behavior by implementing a hierarchical context structure. Level 1 uses rules files for boundaries, Level 2 loads relevant spec sections, and Level 3 pre-loads necessary type definitions and tests to ensure stability.

Rules files function as the foundational layer in context engineering. They operate alongside per-feature specs and selective source data loading to curate agent context precisely for maximizing output quality and consistency.

Compacting context involves summarizing previous progress while safeguarding sensitive information. This practice, combined with surfacing errors succinctly from test results, helps guide debugging and maintains session stability.

Without hierarchical context management, switching tasks causes information loss. Implement persistent rules files for baseline conventions and selectively load per-feature specs to maintain accuracy and consistency across different tasks.